An AI-Made Biden Voice Placed 9,581 Calls Before New Hampshire's January 2024 Primary, and the FCC, the EU and India Have Since Written Different Rules for AI Fakes in Elections

On Sunday 21 January 2024, two days before New Hampshire’s Democratic presidential primary, a recorded voice that sounded like President Joe Biden had gone out on 9,581 phone calls by 7:12 p.m., according to a Federal Communications Commission (FCC) order. The message told listeners to save their vote for the November election instead of voting that Tuesday.

The FCC found that political consultant Steve Kramer arranged the calls and fined him $6 million. A New Hampshire jury later acquitted him of every criminal charge. This piece sets out those outcomes, a documented case from Slovakia’s 2023 election, and what India, the EU and three large platforms now require when AI makes a political fake, with dates as of 8 October 2026.

What the FCC says happened in New Hampshire

The FCC’s forfeiture order says Kramer approached an acquaintance familiar with AI technology about making the message and sent a script on Saturday 20 January. He then asked a robocall company, Voice Broadcasting, to send it. The order quotes the message: “Your vote makes a difference in November, not this Tuesday.”

The caller ID showed a number belonging to the spouse of a former state party chair. The FCC says that person and their spouse had nothing to do with the calls.

The FCC’s ruling and its $6 million fine

The FCC adopted a declaratory ruling on 2 February 2024 and released it on 8 February. It confirms that voice cloning and similar AI tools produce an “artificial” voice under the Telephone Consumer Protection Act, because “a person is not speaking them”. A robocall that uses one therefore needs the called person’s prior consent.

The fine rests on a different law. In an order adopted on 26 September and released on 30 September 2024, the FCC imposed $6,000,000 on Kramer under the Truth in Caller ID Act, which bans misleading caller ID information sent with intent to defraud or cause harm. The amount is $1,000 for each of 3,000 calls the FCC verified, doubled for egregiousness. Commissioner Simington dissented.

Lingo Telecom, the first carrier on the call path, settled its case in a consent decree of 21 August 2024 with a $1 million civil penalty, down from a proposed $2 million, according to the order.

A jury acquitted the consultant on all counts

The criminal case was brought separately in New Hampshire. On 13 June 2025 a Belknap County Superior Court jury acquitted Kramer on all charges, 11 counts of voter suppression and 11 of impersonating a candidate, the Associated Press reported. His lawyers argued that the primary was a straw poll the national party had not sanctioned, so the voter suppression law did not apply, and that the message never named Biden, who was not a declared candidate. Kramer told the AP he would not pay the FCC fine.

Slovakia: a fake interview two days before the vote

Two days before Slovakia’s parliamentary election of 30 September 2023, audio clips circulated that claimed to record a conversation between Michal Šimečka, leader of the Progressive Slovakia party, and a journalist from the daily Denník N about voter fraud. The OSCE’s election office (ODIHR), in its final report of 29 February 2024, calls the clip a notable example of AI-generated manipulative content.

The clips arrived inside the 48-hour campaign silence. ODIHR says that silence binds traditional media and campaign events but not organic, unpaid posts on social networks. Šimečka and the journalist both rejected it as fake.

Whether it changed the result is disputed. Lluis de Nadal and Peter Jančárik, writing in the Harvard Kennedy School Misinformation Review in 2024, call the idea that this was “the first swung by deepfakes” reductive. They point to distrust of institutions and to polls that likely understated support for Robert Fico, who won.

What India’s Election Commission asks of parties

The Election Commission of India (ECI) addresses parties and candidates. For how it identifies voters, see what India’s voter ID system actually is.

On 6 May 2024, during the Lok Sabha election, the ECI directed parties not to publish or circulate deepfake audio or video, to remove such content within three hours of its being brought to their notice, to warn the person responsible, and to take persistent problems to the Grievance Appellate Committee under Rule 3A of the IT Rules. News On AIR, the public broadcaster, reported on 25 October 2025 that the ECI had told parties, candidates and campaign representatives to label AI-altered images, audio and video as “AI-Generated”, “Digitally Enhanced” or “Synthetic Content”, covering at least 10 per cent of the visible display area, or the first 10 per cent of an audio clip, and to name the entity that made it.

On 19 April 2026, with assembly elections under way in Assam, Kerala, Tamil Nadu, Puducherry and West Bengal, the ECI said that since the schedule was announced on 15 March, over 11 thousand social media posts or URLs had been “identified and acted upon including removal of content, FIR, clarifications and rebuttals”. The release does not say how many involved AI. It repeats that misleading or unlawful AI-generated content must be acted on within three hours of coming to a platform’s notice.

India’s IT Rules: advisories first, then a rule in force since 20 February 2026

The IT ministry (MeitY) began with advisories. A Lok Sabha reply of 25 March 2026, published by the Press Information Bureau, lists one dated 26 December 2023 on intermediaries’ due-diligence duties and one dated 15 March 2024 that it says addresses risks from computer resources allowing synthetic creation or modification of text, audio and audiovisual content. The rule itself came with notification G.S.R. 120(E) of 10 February 2026, which MeitY’s FAQ says took effect on 20 February.

The rules define “synthetically generated information” as audio, visual or audio-visual material, made or altered by computer, that appears real and is or is likely to be perceived as indistinguishable from a real person or event. Routine editing, document preparation and accessibility uses are excluded.

Under the new Rule 3(3), a service that lets users create such material must use technical measures to stop unlawful output, including material that falsely portrays a person’s identity, voice, conduct or statement in a way likely to deceive. It must label everything else prominently, with a prefixed audio disclosure for sound, embed permanent metadata or another provenance marker “to the extent technically feasible”, and not let anyone strip the label. Under Rule 4(1A), significant social media intermediaries, the platforms above a user threshold the government sets, must ask uploaders whether content is synthetic, check the answer with technical measures, and label what is confirmed. Under Rule 3(1)(d), the takedown window after a court order or a reasoned government notice is three hours, and the consolidated text shows it replaced 36 hours.

The notified rule sets no label size, unlike the ECI’s advisory for campaign content.

MediaNama reports a further MeitY advisory dated 8 October 2026, signed by Joint Secretary Ajit Kumar, asking social media platforms to apply “enhanced diligence” to false, manipulated or misleading content and to label or contextualise it. MediaNama notes that the advisory lists no penalties and does not mention the February amendment.

What YouTube, Meta and X say they do

YouTube’s post of 14 November 2023 says creators must disclose realistic content made with altered or synthetic media. A label goes in the description panel, and a more prominent label on the video player for sensitive subjects such as elections. Creators who consistently do not disclose may face content removal or suspension from the YouTube Partner Program.

Meta’s post of 6 February 2024 says it will label images when it detects industry-standard AI indicators, namely the C2PA and IPTC metadata standards. For AI-generated video and realistic audio it added a tool for users to disclose them, with penalties if they do not.

X’s rules page, in a copy archived in 2025, bars manipulated or out-of-context media that may cause “widespread confusion on public issues, impact public safety, or cause serious harm”. It lists media showing a real person “fabricated or simulated, especially through use of algorithms or broader artificial intelligence”. Where X cannot reliably tell whether media is misleading, the page says “we may not take action”. Its listed enforcement steps include restricting reach and suspension.

At state level in the US, Connecticut’s AI law, which began binding companies in part on 1 October 2026, is covered in our explainer.

The EU’s labelling duties and their dates

The Commission’s page on its code of practice describes two halves of Article 50 of the AI Act. Providers must mark and make detectable the content their systems generate. Deployers, the organisations that use those systems, must label deepfakes and certain AI-generated text. The Commission published the final code on 10 June 2026. Signing is voluntary, but the Article 50 duties are legal requirements, which the page lists as applicable from 2 August 2026. About 190 organisations had signed by the end of July.

The European Parliament’s release of 7 May 2026 on the deal with the Council says the law “delays the application of watermarking obligations on AI-generated content until 2 December 2026”. The Council recorded adoption of the act, the Digital Omnibus on AI, on 29 June 2026. Parliament’s release does not say whether the delay covers every system or only those already on the market, and the Commission’s page, which lists Article 50 as applicable from 2 August 2026, does not mention it. The new watermarking date in Parliament’s release is 2 December 2026.