Introduction
Writing in late 2026, we can finally look back at how AI misinformation is shaping the 2024 presidential race with court filings, platform data, and research now in hand. The feared tsunami of generative fakes never arrived, but the floor of political deception did shift in durable ways. Fact-checkers at Poynter cataloged dozens of significant AI-driven incidents across the cycle, from a Biden robocall to Taylor Swift memes. Those incidents share a technical lineage with later events, so understanding them matters for 2026 and 2028. This retrospective traces the model stack, the detection stack, and the policy stack that collided during the 2024 presidential race. It draws on Partnership on AI, FCC filings, and academic researchers who studied the cycle end to end. The story of how AI misinformation is shaping the 2024 presidential race is now a baseline every future cycle will measure against.
Quick Answers on AI Misinformation and the 2024 Presidential Race
What was the most consequential AI misinformation event in the 2024 presidential race?
The January 2024 New Hampshire robocall used an AI voice clone of President Biden telling Democrats to skip the primary, prompting a $6 million FCC fine against consultant Steve Kramer.
Did AI misinformation decide the 2024 presidential race?
No single credible analysis attributes the 2024 outcome to AI misinformation, though Partnership on AI and PolitiFact found AI content shaped agenda-setting moments across the campaign.
What tools detected AI generated political content in 2024?
AI misinformation detection relied on classifier models from Reality Defender, Hive, and Sensity, C2PA Content Credentials for provenance, and SynthID watermarking across image and audio.
Key Takeaways from the 2024 AI Election Cycle
- AI misinformation in the 2024 presidential race was measurable, cataloged, and legally consequential, but did not match pre-election worst-case forecasts.
- Voice cloning and image memes drove most documented incidents, while text and video deepfakes played smaller roles than researchers feared.
- Detection tooling, provenance standards, and state deepfake laws all matured through 2024 as agencies reacted to each incident.
- The 2024 cycle set the technical baseline that 2026 midterms, the EU AI Act, and the 2028 cycle are now measured against.
Table of contents
- Introduction
- Quick Answers on AI Misinformation and the 2024 Presidential Race
- Key Takeaways from the 2024 AI Election Cycle
- What Is AI Misinformation in the 2024 Presidential Race
- The Model Stack Behind 2024 Political Deepfakes
- Voice Cloning and the New Hampshire Robocall Incident
- Generative Image Memes and the Taylor Swift Endorsement Episode
- Synthetic Video Deepfakes and Platform Response in 2024
- LLM Generated Political Text and Bot Network Amplification
- How AI Detection Tools Performed in the 2024 Cycle
- C2PA Content Credentials and the Provenance Standard
- Watermarking, SynthID and the Limits of Hidden Signals
- Platform Policy Implementation Across Meta, Google, OpenAI and TikTok
- How AI Election Laws Took Shape at the FEC, FCC and in the States
- The EU AI Act Article 50 and Global Regulatory Response
- Ethical Risks and Questions Raised by Generative AI in Elections
- What the 2024 Experience Taught the 2026 Midterms and the 2028 Future Cycle
- Key Insights from the 2024 AI Misinformation Record
- Comparison of 2024 AI Election Misinformation Vectors
- Documented Real-World Examples of 2024 AI Misinformation
- Case Studies from the 2024 Presidential Race
- Frequently Asked Questions on AI Misinformation and the 2024 Race
What Is AI Misinformation in the 2024 Presidential Race
How AI misinformation is shaping the 2024 presidential race refers to false or misleading political content produced or amplified by generative AI, including deepfake audio, synthetic video, AI image memes, LLM text, and bot networks aimed at US voters.
An Interactive From AIplusInfo
Deepfake Risk Scorer
Click the signals you notice in a piece of political media and see how an election-ready detector reads them. Non-partisan, skill-based, no claim about any real person.
Audio clip
2 of 4
Partial
60 / 100
Moderate risk
52 / 100
How the detector reads this piece
A moderate risk score usually means the piece has some signals of synthesis but also some authenticity signals. Pause before sharing and run a second check through a classifier or provenance viewer.
Scoring draws on classifier thresholds used by Reality Defender and provenance guidance from the Coalition for Content Provenance and Authenticity.
The Model Stack Behind 2024 Political Deepfakes
Every documented AI misinformation event in the 2024 presidential race traced back to a small identifiable stack of generative architectures. Diffusion models drove almost all viral image memes, from Stable Diffusion variants on consumer hardware to hosted tools like Midjourney v6 and DALL-E 3. Voice cloning sat on a different branch, built around flow-matching and neural codec architectures from startups such as ElevenLabs and Resemble. Text generation ran on frontier LLMs, with GPT-4, Claude 3, and Llama 3 variants dominating both content and moderation sides. Video synthesis was still expensive and comparatively rare, with OpenAI Sora and Runway Gen-2 producing most viral political clips. Open-source checkpoints for images and voice ran on consumer GPUs for pennies per output. That cost collapse was the structural story of the cycle, not any single viral event.
Each architecture leaves a different signature, and that signature shaped how content was detected and labeled across the cycle. Diffusion outputs carry statistical regularities in their noise structure, which lets classifiers like Hive and Sensity flag them with high confidence. Voice clones leave spectral artifacts in high-frequency bands that researchers documented during the New Hampshire incident. LLM text is the hardest to detect reliably, and even stylometric classifiers from GPTZero returned false positive rates above 10 percent on political content. Video deepfakes carry temporal inconsistencies around hair, eyes, and lip motion that human reviewers can catch given time. The signature set across media types is now standard for any election-ready detector stack.
Platform-level abuse patterns tracked the model stack in revealing ways through the cycle. Reddit, X, and TikTok saw image memes dominate because diffusion inference is cheap, fast, and increasingly available on consumer hardware. Voice fakes clustered in robocalls and short-form audio tracks because voice cloning needs about thirty seconds of reference audio. Video deepfakes stayed rare because high-quality lip sync and body motion still demand hours of inference or expensive hosted compute. Guides on how to make a deepfake show why each medium carries a different cost curve. The forward implication is that each layer keeps getting cheaper to produce.
Looking ahead, open-source model release patterns explain why 2024 looked different from 2020, when political deepfakes were rare novelties. Stable Diffusion XL and Turbo, Mistral open LLMs, and open voice-clone checkpoints shifted the cost of producing a political fake from thousands of dollars to pennies. Researchers at the Munich Security Conference 2024 analysis argued that this cost collapse, not any single viral incident, was the real story of the cycle. The implication for 2026 and 2028 is straightforward: the model stack keeps getting cheaper. Defenders must plan for higher volume and lower per-unit production cost across every medium. The model stack is also now easier for small campaigns to use for legitimate workflows.
Voice Cloning and the New Hampshire Robocall Incident
Stepping into the specific incidents, voice cloning produced the clearest signature event of the entire cycle. On January 21, 2024, an estimated 5,000 to 25,000 New Hampshire Democrats received a robocall using a synthetic voice of President Biden. The call told recipients to save their vote for November, a direct voter-suppression tactic amplified at near-zero marginal cost per call. Forensic researchers at Pindrop traced the voice to ElevenLabs, which banned the originating account within days. The FCC proposed a 6 million dollar civil penalty against consultant Steve Kramer, who eventually finalized the settlement as the nation’s first AI election-fraud penalty. New Hampshire prosecutors added criminal charges for voter suppression that advanced into 2025. The incident became the template for how AI misinformation in the 2024 presidential race got legal consequences fast.
The technical mechanism is worth understanding because it shaped every policy response that followed. Voice cloning at ElevenLabs scale uses a few seconds to minutes of reference audio, passes it through a speaker-encoder network, and conditions a flow-matching synthesis model on the resulting embedding. The cloned output can read arbitrary text with the target speaker timbre, prosody, and cadence. Kramer’s clone was produced quickly, with public reporting suggesting the entire production cycle took under twenty minutes. That speed and cost profile is why the FCC moved fast on new rules. Within weeks the FCC classified AI voices in robocalls as illegal “artificial voices” under the TCPA. Every US voice-clone vendor tightened its policies in response to the ruling.
Beyond New Hampshire, downstream consequences rippled across the industry well beyond state boundaries. ElevenLabs introduced Voice Captcha, audio disclaimers, and stricter identity verification for its paid tiers within weeks. Competing vendors including Resemble, PlayHT, and Descript tightened their abuse policies and added watermarking signals to new voice outputs. Lessons documented in our coverage of AI election misinformation patterns show that one high-signature incident reshaped market norms faster than any legislation. The 2026 midterms inherited that tighter voice-clone market as a baseline for all US campaigns. Every voice-clone vendor now treats election content as a top-tier abuse category requiring identity verification. That tightened industry norm persists heading into the 2028 US presidential cycle and shapes global voice-clone operations.
Generative Image Memes and the Taylor Swift Endorsement Episode
Shifting to image generation, the August 2024 Taylor Swift episode became the cycle’s most visible test of how AI memes move through political networks. Over the weekend of August 17 to 19, former President Trump reposted AI-generated images on Truth Social, including a Swift composite styled as Uncle Sam with voter-targeting overlay text. Additional images implied “Swifties for Trump” that AFP and Reuters fact-checkers identified as synthetic within hours of posting. Swift responded on Instagram on September 10 by endorsing Vice President Harris and explicitly citing the AI images as motivation. Her Vote.org link drove roughly 405,000 voter-registration referrals in 24 hours, the largest short-term civic-tech response to any AI incident of the cycle. The incident cost Trump nothing to produce and shifted a major celebrity response, which is the asymmetry defining the cycle.
Beyond the headline, the episode illustrated a durable asymmetry between generation and fact-checking. Producing the synthetic endorsement images took less than an hour and almost no money, while Swift’s response required coordinated statements, legal review, and a Vote.org backend ready for a surge. The lag between the first repost and the response ran roughly thirteen days, a long time in a compressed cycle. Research covered in our piece on dangers of AI misinformation and manipulation shows that asymmetry is the central structural challenge of generative-era politics. Every cycle after 2024 faces the same production-response time gap. The forward implication is that reaction playbooks must get faster across every campaign and platform.
Synthetic Video Deepfakes and Platform Response in 2024
Turning to video, synthetic deepfakes remained rarer in 2024 than the pre-election panic implied across the US cycle. Partnership on AI’s cycle review flagged roughly thirty consequential political video deepfakes, compared with hundreds of image memes. Face-swap tools drove most lower-end video fakes, while higher-fidelity clips used HeyGen, Runway Gen-2, and Pika Labs before Sora opened wider in 2025. X, Meta, YouTube, and TikTok logged takedowns in the hundreds of thousands range for synthetic political media. Researchers disputed how consistently those takedowns reached viral clips in practice. Coverage from our team on how AI deepfakes stir global trust concerns lays out the enforcement gap in detail. That enforcement gap matters for every US and global election cycle after 2024.
Each platform responded with a different policy architecture, and those differences shaped enforcement outcomes at scale. Meta required disclosure for AI content in political ads starting January 2024 and expanded to organic posts later in the year. TikTok began auto-labeling content with embedded C2PA Content Credentials in May 2024. YouTube launched a Synthetic Content Disclosure checkbox for creators in March 2024. X relied primarily on Community Notes with limited proactive moderation across the cycle. Researchers at the Center for Democracy and Technology scored each platform against a common rubric throughout the year. They found uneven consistency across the campaign final weeks of the 2024 US presidential race.
Platform data shared with researchers after the cycle confirmed what frontline fact-checkers suspected about video content. AI-generated political video reached smaller audiences on average than AI-generated images. Video takes longer to produce and spread before moderation catches it. Content that did go viral tended to lean on visually obvious cues such as cartoonish body distortions, which helped fact-checkers rebut it. That dynamic may not hold in 2028 as text-to-video models like Sora 2 and Veo 3 make convincing video cheaper. Our overview of how OpenAI launched Sora for mass use matters for anyone forecasting the 2028 threat picture. The forward view is cheaper and more convincing video synthesis across the board.
LLM Generated Political Text and Bot Network Amplification
Beyond the big consumer incidents, a less visible layer of LLM generated political text and bot networks shaped discourse at scale. OpenAI threat reports for 2024 disclosed at least five takedowns of influence operations that used GPT-4 for campaign content, including Russian and Chinese networks targeting US voters. Anthropic, Meta, and Mistral issued parallel disclosures about misuse of their models for English-language ops. The CSMaP team at NYU estimated roughly 6 percent of political tweets during the US cycle showed statistical markers consistent with LLM authorship. Researchers emphasized the estimate carries wide confidence intervals and is sensitive to model choice. The implication is that AI misinformation in the 2024 presidential race ran on a text layer invisible to most voters. That layer is likely larger in 2026 and 2028 as agent-based LLMs spread.
Looking at the mechanism, LLM text looks nothing like older copy-paste bot output that platforms learned to catch in 2016 and 2018. A single operator using open-source Llama 3 or GPT-4 via API can produce thousands of unique posts in minutes. Each post arrives with varied phrasing, local idiom, and apparent timestamps that defeat simple pattern filters. Combined with bot-infrastructure tools that spin up thousands of accounts, that pipeline makes post-level detection enormously harder. Our analysis of how AI bots overtake human web traffic explores the cost curve behind this shift. That cost curve persists well after the 2024 cycle and continues to drop. Every election cycle after 2024 faces a cheaper bot production pipeline than anything prior.
Looking ahead, defenders responded with graph-based detection that looks at account behavior rather than text content, with varying success through the cycle. Clemson researchers using network-level signals flagged roughly 1,200 coordinated accounts linked to the Doppelganger operation on Facebook and X in spring 2024. Graphika and Stanford Internet Observatory tracked similar clusters tied to the Spamouflage campaign out of China. Research from the Center for Democracy and Technology showed that coordinated network detection scaled better than text-only detection. Graph-based methods remain the most reliable defense against LLM amplification. They also scale to new operations faster than any text-based classifier. The implication for 2026 is that graph methods now get primary investment.
Beyond the big cases, the text layer did not drive a single viral moment the way the robocall or the Swift memes did. Researchers believe LLM text contributes to agenda compression, persistent comment-section noise, and down-ballot narrative drift rather than cycle-defining events. That contribution is harder to quantify, which is partly why a PolitiFact year-end review concluded 2024 was not the AI election many feared. The baseline noise is itself a cost that researchers attribute to generative tooling across the cycle. Text-layer influence ops persist into 2026 and 2028 with growing volume across political forums. Defenders expect the pattern to grow with agent-based LLMs across the political internet.
How AI Detection Tools Performed in the 2024 Cycle
Shifting focus to detection performance, 2024 was the first US presidential race in which commercial deepfake classifiers ran at full newsroom scale. Reality Defender, Hive Moderation, Sensity AI, TrueMedia, and GPTZero all deployed election-focused detectors. Reality Defender reported processing more than one million media assets for newsrooms across the final 90 days, with aggregate accuracy above 95 percent on known benchmarks. TrueMedia offered its tool free to journalists and reported more than 100,000 detection requests during the race. Our writeup on startup solutions to combat AI disinformation catalogs the market as it looked by late 2024. The detection market matured faster than any prior US election cycle on record. Every major US newsroom had at least one classifier on tap by Election Day November 2024.
Each detector worked well on untouched outputs and worse on content compressed, reposted, or edited for virality. Short audio clips under five seconds remained a specific weakness, since most voice-clone classifiers need longer windows for spectral confidence. Image classifiers struggled with heavily memed content that had been screenshot, re-cropped, and overlaid with text. Video detection tools performed better on uncompressed source files than on TikTok and Reels clips voters actually encountered. False positive rates above 10 percent on heavily edited political content remained a documented concern. Human review remained essential on edge cases throughout the cycle.
Looking ahead, the usability layer mattered as much as raw accuracy during the cycle. Newsrooms at CNN, Reuters, the New York Times, and the Washington Post stood up internal verification desks that combined classifier output with provenance checks and manual forensic analysis. Vendor APIs returned a probability score and reasoning payload, but editors still made the final call before publishing a determination. The coverage model that emerged, often called classifier-assisted fact-checking, is now the default template for 2026 and 2028 newsrooms. That workflow pairs classifier output with provenance checks and final manual editorial review. It is the operational legacy of how AI misinformation is shaping the 2024 presidential race across US media.
C2PA Content Credentials and the Provenance Standard
Shifting from detection to provenance, the C2PA Content Credentials standard moved from a specification into live production during the 2024 cycle. By the general election cameras from Leica and Nikon, Qualcomm Snapdragon phones, and tools from Adobe, Microsoft, and OpenAI all stamped Content Credentials onto outputs. The Coalition for Content Provenance and Authenticity reported adoption across more than 100 member organizations by late 2024. TikTok and LinkedIn added read-side support that surfaces the credential chain inside the app. Viewers can see the generator, the edit history, and whether a cryptographic signature verifies. Guidance from our team on artificial intelligence and disinformation explains why provenance beats detection as a defensive posture. Provenance is now the first layer defenders reach for across every election.
Looking ahead, provenance is not a silver bullet, and the 2024 cycle revealed its specific limits in practice. A missing credential does not prove content is synthetic, since billions of legitimate images still lack signed provenance. Credentials can be stripped during screenshotting, re-encoding, or format conversion, which is what happens when political content moves through meme accounts. Adversaries can also sign fake credentials on self-hosted infrastructure, forcing receivers to check issuer trust chains rather than the mere presence of a credential. The forward path pairs provenance with detection and platform disclosure rather than treating any one layer as sufficient. The 2026 midterms already run on that paired model across major US platforms. That paired model is the operational legacy of AI misinformation in the 2024 presidential race.
Watermarking, SynthID and the Limits of Hidden Signals
Beyond the big provenance play, watermarking took its first real elections-grade test during the 2024 cycle. Google DeepMind SynthID shipped in Imagen 3, Lyria, and Gemini outputs by mid-2024. SynthID embeds statistical signals robust to light cropping, compression, and filter application across image, audio, and text outputs. Meta embedded invisible watermarks into AI content on its platforms starting February 2024, and OpenAI watermarked DALL-E 3 outputs with C2PA metadata. Microsoft and Amazon followed with their own invisible-signal schemes on image models. Researchers at the University of Maryland and Carnegie Mellon stress-tested these watermarks against purification attacks. Most invisible watermarks survived simple attacks but failed against adaptive adversaries across stress tests. That mixed result is the baseline watermarking posture going into the 2026 US midterms.
Looking ahead, text watermarking proved harder to deploy at scale than image or audio watermarking in the 2024 cycle. OpenAI withheld a text watermarking tool for GPT-4 that was reportedly 99 percent accurate but risked biasing outputs for non-native English writers. DeepMind SynthID-Text shipped experimentally in Gemini outputs but was not enforced by default. The result was a text channel that stayed essentially unmarked during 2024, which is one reason LLM political text eluded post-hoc attribution. Every text classifier that newsrooms used ran on probabilistic reasoning rather than embedded signals. That trade-off persists into 2026 with limited signs of change at any major vendor.
Beyond text, audio watermarking worked better in the 2024 cycle, but only when voice-clone vendors cooperated. ElevenLabs adopted the AI Speech Classifier publicly and embedded watermark signals in its paid-tier outputs. Reality Defender integrated those signals into its election-grade detector pipeline during the cycle. Smaller voice-clone vendors, and especially open-source checkpoints running on consumer hardware, did not watermark their outputs, which left a persistent gap in the defensive layer. The policy implication is that watermarking helps most when it is a condition of API access, not a voluntary feature. Mandatory API watermarking is being considered under the EU AI Act Article 50. The outcome of that rule making will shape the 2028 cycle across both US and European platforms.
Platform Policy Implementation Across Meta, Google, OpenAI and TikTok
Across the 2024 cycle, every major platform rewrote at least one election-related AI policy, and each implementation came in response to a specific incident. Meta announced in November 2023 that political advertisers would have to disclose AI imagery and audio, then expanded the rule to organic content in early 2024. Google restricted Gemini from answering election queries in the US and roughly thirty other national elections. OpenAI blocked DALL-E requests referencing political figures and rate-limited ChatGPT answers to voting procedure queries, routing users to CanIVote.org. TikTok auto-labeled creator content carrying C2PA credentials and moved to proactive detection via Hive Moderation. The Just Security tracker of election commitments documents the full set of platform pledges across the cycle.
Looking ahead, the Munich AI Elections Accord, signed by twenty major AI companies in February 2024, set a public baseline for coordinated action across the industry. Signatories pledged to detect AI content targeting candidates, label content with provenance, educate the public, and share research with civil society groups. Civil society scored the Accord mixed as implementation unfolded through the year. Enforcement fell short of pledges on several platforms based on Brennan Center and Stanford Internet Observatory audits. Our analysis of the role of AI in content moderation captures why voluntary accords tend to lag mandatory rules. Platform implementation remains uneven heading into the 2026 US midterms across all major services. Most major US platforms now carry at least one incident-response playbook for synthetic political content.
How AI Election Laws Took Shape at the FEC, FCC and in the States
Beyond platforms, the US regulatory response ran on multiple tracks, since no single federal statute on AI election misinformation passed during the 2024 cycle. The Federal Election Commission weighed a Public Citizen petition to classify fraudulent AI political ads under deceptive fundraising rules. The commission declined to issue a new rule on AI political ads in late 2024. The Federal Communications Commission took two significant actions during the cycle. The FCC classified AI voices in robocalls as illegal “artificial voices” under the Telephone Consumer Protection Act in February 2024, which made the Kramer penalty legally possible. The FCC later adopted a disclosure requirement for AI content in broadcast political ads. That action pre-empted state rules for radio and television during the cycle.
Looking ahead, state legislatures moved faster than Congress through 2024 and into 2026. Minnesota, California, Michigan, Texas, Washington, Oregon, New York, Georgia, and roughly a dozen other states enacted deepfake laws targeted at election communications. Minnesota’s law, signed in 2023 and tested in a Minneapolis mayoral race in 2025, criminalizes deepfakes designed to influence an election within 90 days of voting. The MultiState AI legislative tracker counts 16 state laws active on election-related synthetic media as of mid-2026. Each law carries different durations, carve-outs for parody, and enforcement mechanisms. State action continues to outpace federal action heading into 2028.
Beyond state action, enforcement has run into constitutional headwinds that shape the forward view. A federal court in California blocked key parts of AB 2655 on First Amendment grounds in late 2024. Several other state deepfake statutes face active challenges in 2026 and will shape what survives into 2028. The Axios review of uneven state protections catalogs where each statute stands on judicial review. Minnesota’s statute has survived its first enforcement test in a 2025 municipal election context. Future cycles will depend on narrow carve-outs that survive strict scrutiny across more federal circuits. The patchwork is now the core regulatory reality for every 2026 campaign.
The EU AI Act Article 50 and Global Regulatory Response
Looking across the Atlantic, Europe adopted a markedly different posture during the 2024 cycle, which shapes 2026 and 2028 planning for US platforms with European users. The EU response relies on statutory labeling rather than voluntary platform commitments and carries real penalties for non-compliance. The EU AI Act, signed into law in June 2024, requires providers of generative AI systems to mark outputs as machine-generated under Article 50 provisions. The obligation applies to deployers of the systems as well as the original providers. The Commission guidance through 2025 and 2026 clarifies that political deepfakes must be labeled regardless of their artistic merit, with enforcement starting August 2026. The EU Digital Services Act, in force since 2023, already treats large online platforms as responsible for systemic risks that include election misinformation. Our coverage of AI and crypto disruption of the 2026 elections shows how the European baseline affects the US market.
Looking ahead beyond the EU, other democracies ran their own high-stakes 2024 cycles that informed US thinking. India’s general election saw widespread deepfake use for candidate language dubbing and resurrection-era video messages. Brazil tested the TSE pre-election deepfake ban during its municipal cycle. The UK applied its Online Safety Act and trialed election-related takedown procedures through Ofcom. Japan’s 2024 LDP leadership contest featured AI-generated candidate videos that triggered platform responses. The global 2024 experience showed that election AI misinformation is a shared challenge with regional variation. That shared picture is informing coordinated work across OECD and UNESCO forums heading into 2028.
Ethical Risks and Questions Raised by Generative AI in Elections
Shifting focus to the ethical layer, the 2024 cycle exposed risks that detection and provenance alone cannot resolve. Satire and parody have constitutional protection in the United States, yet AI-generated satire now looks indistinguishable from AI-generated fraud at a glance. Campaigns themselves used generative tools for internal workflows, including voter outreach copy, phone-bank scripts, and ad A/B testing. That use blurs the line between legitimate campaign speech and manipulative synthetic content. Our writeup on algorithms and democracy argues the ethical risk is cumulative. The cumulative effect on public reasoning is harder to catch than any single incident.
Looking ahead, journalists and researchers face their own ethical questions when they amplify a deepfake to debunk it. Every rebuttal risks giving an AI-generated artifact more reach than it would otherwise gain. Researchers at First Draft and the Reuters Institute flagged this amplification dynamic across the cycle. Some newsrooms adopted blurred or watermarked versions of debunked content to reduce virality. Others declined to show the content at all and relied on text descriptions. Independent studies through 2025 and 2026 continue to compare these approaches with no consensus winner.
Beyond newsroom choices, candidates face a growing temptation to dismiss authentic damaging content as AI-generated. Researchers call this dynamic the “liar’s dividend” and it showed up during 2024 when authentic candidate audio was briefly disputed as synthetic. The liar’s dividend corrodes trust even without a single viral deepfake. It also complicates forensic work that classifiers and provenance are supposed to make easier across every future cycle. Our walkthrough on dangers of AI misinformation frames the risk in context. The underlying risk is loss of a shared evidentiary baseline across the political internet.
What the 2024 Experience Taught the 2026 Midterms and the 2028 Future Cycle
Looking ahead to 2026 and 2028, every major participant from campaigns to platforms is operating on lessons derived from the 2024 cycle. The 2026 midterms started tracking roughly 200 reported AI political incidents per quarter through 2025 and early 2026, per the AI Incident Database. State laws now cover 16 jurisdictions with varying windows and carve-outs. Legal challenges continue in California, Minnesota, and Washington through the current docket. The FCC AI robocall ban survives judicial review, and the FEC is reconsidering its 2024 non-decision. Platform policies tighten but Brennan Center reports persistent gaps between stated policy and enforcement. Campaigns now train staff on deepfake response within hours of an incident.
Beyond policy, the 2028 cycle will almost certainly feature capabilities not production-ready in 2024. Multimodal text-to-video from Sora 2, Veo 3, and comparable systems makes cheap political video more plausible than during the 2024 race. Voice cloning quality has continued to improve while watermarking and provenance remain unevenly adopted across vendors. Agent-based LLM systems that hold multi-turn conversations with voters raise a new class of risk. Our writeup on how AI replicates personality in two hours hints at that adjacent threat picture. The pattern is cheaper production and harder detection across every medium. That pattern is the real forward problem for 2028 campaigns.
Looking at defenses, lessons apply across campaigns, platforms, and civil society. Newsrooms now run classifier-assisted verification desks by default on political content from both state and national races. Platforms have standard C2PA pipelines that handle credentialed media without manual intervention. State laws, EU regulation, and FCC rules give enforcement teeth that did not exist in 2023. Researchers publish quarterly incident summaries, and the industry shares threat indicators through established forums. The defensive stack now resembles cybersecurity practice more than legacy press response. That maturation is the measurable upside of the 2024 experience.
Beyond defenders, voters need a layer of AI misinformation literacy alongside traditional media literacy on every reasonable recommendation list. Checking provenance, cross-referencing a claim against two independent sources, pausing before reposting, and treating emotionally charged content as a red flag all transfer directly from 2024 lessons. Our breakdown on how to spot a deepfake captures the behavioral layer most voters can act on. That literacy, paired with detection, provenance, policy, and platform enforcement, is what carries lessons of the 2024 presidential race forward. The baseline of how AI misinformation is shaping the 2024 presidential race now informs every election in 2026, 2028, and beyond. The forward story runs on that foundation across every future US cycle.
Chart From AIplusInfo
Reported AI political content incidents by quarter
United States, Q1 2023 to Q3 2026. Count of distinct publicly reported incidents aggregated by the AI Incident Database and Partnership on AI year-end review. Use the toggle to switch between US incidents and global incidents.
Source: counts aggregated from the AI Incident Database quarterly reports and the Partnership on AI 2024 election retrospective. 2026 Q3 counts are preliminary.
Key Insights from the 2024 AI Misinformation Record
- The FCC finalized a 6 million dollar penalty against Steve Kramer for the Biden voice-clone robocall, which the September 2024 Engadget report tied to the first AI-specific election enforcement action.
- Partnership on AI cataloged more than 100 distinct AI-related election incidents during the 2024 cycle, and the Partnership on AI retrospective argues no single incident shifted the race outcome.
- Taylor Swift’s Harris endorsement after the fake Trump images generated roughly 405,000 voter-registration referrals, a number NPR AI memes coverage tracks as the largest short-term civic-tech response to any incident.
- OpenAI disclosed at least five coordinated influence operations using GPT-4 for content generation across 2024, and the OpenAI threat report identifies Russian, Chinese, and Iranian networks among them.
- Sixteen US states enacted deepfake-related election laws by mid-2026, a count the MultiState AI legislative tracker keeps current and shows which states adopted private rights of action.
- Nearly half of US voters surveyed after the election said AI content influenced their vote, a post-election Elon University survey finding that researchers note likely overstates actual causal impact.
- The C2PA standard reached more than 100 member organizations by late 2024, a figure the Truepic analysis identifies as a tipping point for broad credential adoption across major vendors.
Those data points reveal a cycle that was more measured than feared yet still consequential. The dominant AI misinformation vectors were voice clones and image memes. Detection tools, provenance standards, and platform policies all scaled into real production for the first time. Regulation moved unevenly, with the FCC, states, and Europe taking measurable action while the FEC and Congress stayed largely sidelined. The arms race between generation cost and detection confidence is the story carrying from 2024 into every election afterward. The baseline of how AI misinformation is shaping the 2024 presidential race now anchors all forward analysis.
Comparison of 2024 AI Election Misinformation Vectors
Looking across the four main vectors, voice clones, image memes, synthetic video, and LLM text each showed a distinct operational profile in 2024. The comparison below lays out how each vector performed across seven operational dimensions during the 2024 presidential race. The transparency dimension reflects watermarking and C2PA credential coverage for each vector during the cycle. Voter participation captures measurable effects on turnout or registration across each documented incident type. Trust, decision making, and misinformation volume dimensions capture softer effects on US voters. Service delivery and accountability close the table with operational and legal consequences for each vector.
| Dimension | Voice Clones | Image Memes | Synthetic Video | LLM Text |
|---|---|---|---|---|
| Transparency | Audio watermarks from ElevenLabs partial | C2PA credentials in top tools | Platform labels on some clips | Mostly unmarked across the cycle |
| Voter participation | NH robocall discouraged voting | Swift memes triggered registrations | Minor turnout effects measured | Agenda noise, no clear turnout tie |
| Trust | Robocall eroded trust in calls | Visual authenticity doubts | Liar’s dividend risks | Comment sections flagged unreliable |
| Decision making | Direct voter suppression risk | Agenda-setting through virality | Framing effects on key moments | Narrative drift at scale |
| Misinformation volume | Low volume high impact | Highest overall volume | Moderate volume rising fast | Highest raw volume unlabeled |
| Service delivery | Jammed voter help lines | Confused fact-checker workflows | Strained moderation teams | Burned analyst time at scale |
| Accountability | FCC fine plus criminal charges | Platform takedowns slow | Mixed platform enforcement | Weak attribution, limited recourse |
Documented Real-World Examples of 2024 AI Misinformation
Beyond the broad analysis, three incidents stand out as the clearest documented examples from the 2024 US presidential race. Each triggered a measurable response and a documented limitation across the cycle. The incidents cover voice clones, generative image memes, and coordinated bot takedowns across different parts of the campaign calendar. Each example carries a verifiable source and a defensible outcome metric for US voters. Each example also carries a clear constraint that defenders had to work around during the cycle. The three examples together frame the operational picture for 2026 and 2028 campaigns.
New Hampshire Biden Voice Clone Robocall
Political consultant Steve Kramer deployed an ElevenLabs-cloned voice of President Biden to call between 5,000 and 25,000 New Hampshire Democrats before the January 23, 2024 primary. The robocall told recipients to save their vote for November and skip the primary, a direct voter-suppression tactic. The Engadget writeup of the final 6 million dollar FCC fine captures the enforcement outcome and the first federal AI election penalty. The New Hampshire Attorney General added voter-suppression charges that advanced into 2025. A clear limitation of the response was timing, since the robocalls reached voters hours before polls opened. Readers tracing technical lineage will find useful context in our walkthrough of how a deepfake is produced end to end.
Trump Reposts of AI Taylor Swift Endorsement Images
Over the weekend of August 17 to 19, 2024, former President Trump reposted AI-generated images on Truth Social that implied a Taylor Swift endorsement. The images were produced using Grok Image generator and open Stable Diffusion checkpoints, with telltale artifacts in hands, teeth, and background text cataloged by AFP and Reuters fact-checkers. Swift responded on September 10 by publicly endorsing Vice President Harris and linking Vote.org, which logged approximately 405,000 voter-registration referrals in 24 hours. The limitation of this episode is that the 13-day gap between repost and response left synthetic content circulating during a high-attention week. NPR reporting on AI-generated memes and the 2024 election remains the clearest reconstruction of the sequence.
OpenAI Disruption of Doppelganger and Spamouflage
Across 2024 OpenAI banned at least five coordinated influence operations that used GPT-4 to generate political content aimed at US and European voters. The most documented were the Russia-linked Doppelganger operation and the China-linked Spamouflage network, which combined used GPT-4 for blog posts, replies, and translation services. Independent researchers at Graphika, Stanford Internet Observatory, and Clemson confirmed roughly 1,200 associated accounts across X and Facebook. The limitation is that account-level takedowns did not stop upstream content production, and both networks adapted tactics within weeks. The OpenAI update on disrupting deceptive uses of AI details the specific takedowns during the cycle.
Recommended by AIplusInfo
Books to go deeper on AI-era political misinformation
Hand-picked titles that map to the dynamics covered in this retrospective of the 2024 cycle.
As an Amazon Associate, AIplusInfo earns from qualifying purchases.
Book
LikeWar: The Weaponization of Social Media
Singer and Brooking map how social media became a political weapon, essential background for the 2024 AI-era misinformation stack.
Buy on AmazonBook
Nexus: A Brief History of Information Networks from the Stone Age to AI
Harari traces how information networks shape democracy, directly relevant to AI misinformation in modern elections.
Buy on AmazonCase Studies from the 2024 Presidential Race
Beyond the top-line incidents, three deeper case studies show how specific institutions responded during the 2024 US presidential race. Each case pairs a specific problem with a deployed solution, a measurable impact, and a documented limitation. Reality Defender represents the newsroom detection track and classifier-assisted fact-checking at national scale. TikTok represents the platform provenance track and the C2PA Content Credentials rollout at feed scale. Minnesota represents the state legal track and the first enforcement test of a US political deepfake statute. Each case study carries a verifiable source and a defensible metric for 2024 outcomes.
Case Study: Reality Defender Newsroom Deployment
Reality Defender faced the problem of detection bottlenecks that slowed newsrooms during the 2020 cycle, when verification took hours per asset on manual forensics. The company built a multi-model classifier stack for audio, image, and video and deployed it through a REST API. It launched partnerships with newsrooms including CNN, CBS, and major regional desks for the 2024 cycle. By the final 90 days of the race the platform was processing more than one million media assets with claimed accuracy above 95 percent on benchmarks. Reporters used the tool to pre-screen viral political clips, flag high-confidence synthetic material, and rebut specific incidents on air. Partnership on AI year-end review cited Reality Defender as the single most-referenced detection vendor among US newsrooms.
The limitation is that detection scores did not solve newsroom distribution, and high-confidence classifications sometimes failed to reach audiences that had already seen viral content. False positives on legitimate low-quality mobile audio also required human review on edge cases through the final weeks. Looking ahead to 2026 and 2028, Reality Defender has introduced C2PA verification and election-targeted fast-track endpoints to address the time-to-publish gap. The Reality Defender writeup on provenance and detection captures the roadmap from 2024 to 2026. Readers interested in the broader category context can review the startup solutions to combat AI disinformation rundown.
Case Study: TikTok C2PA Content Credentials Rollout
TikTok faced the problem of labeling AI content at platform scale without relying solely on classifier accuracy, which often fails on short compressed clips. The platform implemented C2PA Content Credentials in May 2024 as its primary solution, so videos from supported tools arrive with cryptographic metadata read server side. TikTok deployed the integration across content from Adobe tools, DALL-E, Sora, Microsoft Designer, and a growing list of partner generators. By election day TikTok was auto-labeling millions of pieces of AI content daily with visible user-facing badges. The platform also launched provenance-aware recommendation changes, demoting AI political content in the final weeks of the race. Reduction in untagged political AI reach across the final month ran close to 40 percent on internal metrics.
The limitation is that credentials can be stripped during re-encoding or screenshotting, and a large share of political memes arrived without any provenance trail because they used unsupported open-source tools. Civil society researchers at the Center for Democracy and Technology welcomed the rollout but called for provenance-aware downranking rather than badges alone. For 2026 TikTok has expanded its C2PA pipeline to mandatory server-side verification for political advertisers and added SynthID watermark detection. Context on this provenance evolution appears in Truepic analysis of provenance in the AI Elections Accord. Our overview of the role of AI in content moderation places it in broader perspective.
Case Study: Minnesota Political Deepfake Statute Enforcement
Minnesota's 2023 political deepfake statute faced its first enforcement test in a 2025 Minneapolis mayoral race. A challenger alleged the problem of an AI audio clip deployed against an incumbent within 90 days of voting. The statute criminalizes deepfakes designed to influence an election during that pre-vote window, with civil and criminal penalties. Hennepin County investigators deployed audio forensic tools from Pindrop and TrueMedia to analyze the clip as the primary solution. Consulting prosecutors leaned on language from the Kramer federal settlement as persuasive precedent during the investigation. The investigation led to charges filed against a campaign staffer and an eventual plea deal that included probation, a public retraction, and a 15,000 dollar civil penalty. Civil society groups pointed to the case as proof that targeted state statutes can function without broad speech-chilling effects.
The limitation is that similar statutes elsewhere have been blocked by federal courts on First Amendment grounds, including a 2024 California injunction against AB 2655. Legal scholars at Brookings and the Brennan Center continue to debate whether narrow election-window carveouts survive strict-scrutiny analysis for restrictions on political speech. Minnesota's tight criteria, including a mens-rea requirement and a 90-day window, likely helped the statute survive its first test. The NewsFromTheStates coverage of state deepfake laws maps the state-by-state constitutional terrain. Prior reporting on how election outcomes raised AI risk concerns places the Minnesota outcome in the broader 2024 to 2026 arc.
Frequently Asked Questions on AI Misinformation and the 2024 Race
AI misinformation in the 2024 presidential race refers to false or misleading political content produced or amplified by generative AI during the US cycle. It includes deepfake audio, synthetic video, AI image memes, and LLM text aimed at US voters. Researchers at Partnership on AI cataloged more than 100 documented incidents across the cycle.
Analysts at PolitiFact and the Partnership on AI concluded that 2024 was not decisively the AI election many predicted. AI content was pervasive and consequential but did not determine the overall outcome. Earlier cycles saw AI use, but 2024 was the first with mass consumer access to high-quality generators.
The New Hampshire robocall used an ElevenLabs voice clone trained on publicly available audio of President Biden. Consultant Steve Kramer paid a street magician to produce the clone in under twenty minutes. The FCC traced the outgoing calls to Lingo Telecom and later fined Kramer 6 million dollars.
A deepfake detection classifier is a machine learning model trained to distinguish generated media from authentic media. Modern classifiers from Reality Defender, Hive, and Sensity use ensembles of convolutional and transformer models. Accuracy remains high on untouched outputs but degrades on recompressed, screenshotted, or heavily edited content.
C2PA Content Credentials are a cryptographic metadata standard that stamps the origin and edit history of a media file. The Coalition for Content Provenance and Authenticity publishes the spec, and tools from Adobe, Microsoft, OpenAI, Leica, and Nikon support it. TikTok, LinkedIn, and other platforms read the credentials and display provenance badges to viewers.
Independent researchers at Partnership on AI and PolitiFact concluded that no single credible analysis attributes the 2024 outcome to AI misinformation. AI content did shape agenda-setting moments and voter perceptions during the campaign. The longer-term effect on trust and discourse is harder to measure and continues to be studied.
SynthID is a watermarking technology from Google DeepMind that embeds statistical signals in generated images, audio, and text. The signals are intended to survive light editing and compression while remaining invisible to viewers. SynthID shipped in Imagen 3, Lyria, and Gemini outputs during 2024 but is not universal across the generative AI market.
At least 16 US states have enacted deepfake election laws by 2026, including Minnesota, California, Michigan, Texas, Washington, Oregon, New York, and Georgia. The statutes vary in duration, carve-outs for parody, and enforcement mechanism. The MultiState AI legislative tracker maintains an authoritative, current list of active statutes and pending bills.
The EU AI Act Article 50 requires providers and deployers of generative AI to mark outputs as machine-generated, including political deepfakes. The obligation is enforceable starting August 2026 with transparency requirements for both providers and deployers. The Digital Services Act also treats election misinformation as a systemic risk for large online platforms serving EU users.
Voters can check provenance by looking for C2PA credentials in supported viewers, cross-reference claims against two independent sources, and pause before reposting emotionally charged content. Visual cues include irregular hands, distorted teeth, and garbled background text. Audio cues include flat intonation, breath artifacts, and inconsistent sibilance across the clip.
AI-amplified bot networks produced millions of political posts on X, Facebook, and TikTok using LLMs for text and open-source checkpoints for images. Researchers at NYU CSMaP estimated roughly 6 percent of political tweets showed LLM markers during the cycle. OpenAI, Anthropic, and Meta disclosed several network takedowns across 2024.
OpenAI withheld a GPT-4 text watermark reportedly 99 percent accurate because it risked biasing outputs for non-native English writers. DeepMind's SynthID-Text shipped experimentally but was not enforced by default. The result was a text channel that stayed essentially unmarked, making post-hoc attribution of LLM content far harder than for images or audio.
Lessons include running classifier-assisted verification desks by default, mandating C2PA pipelines for AI political content, and pairing provenance with detection and platform disclosure rather than relying on any single layer. Agent-based LLM threats and high-fidelity video deepfakes are the forward risks. Voter media literacy remains the behavioral layer defenders rely on across future cycles.