Search Ctrl + K

First thoughts · Requirements · AI and the profession

Whether AI is superseding voiceover artists

Voiceover artist is the right term, and the profession is not being superseded. It is being cut in half. Synthetic voices have taken the work that was bought by the yard, and they have taken it fast. In the National Association of Voice Actors survey of January and February 2026, 21 per cent of 1,379 respondents said they had knowingly lost a job to a synthetic voice, against 14 per cent in the same survey a year earlier. Over the same period 41 per cent of those respondents said their income had risen and 30 per cent said it had fallen.

The line falls between work where the buyer wants a voice and work where the buyer wants a person, not between human and machine. Where the client is buying audio to fill a slot, the machine has already won on price and will not give the ground back. Where the client is buying judgement, accountability, a name on an invoice or a performance that can be redirected in the room, the machine has made almost no progress, and the buyers say so.

This matters for Anita in a specific way, set out at the end: the segments she has named are among the better-defended ones, but the route new entrants used to take into the profession has closed.

What has actually been measured

The NAVA survey is the best regular count of the profession, and it should be read with its limits in view. Its 1,379 respondents were recruited through social media, voiceover Facebook groups and a contact list of roughly 4,000 people, so they selected themselves and skew towards the United States and towards people already engaged with professional bodies. Seventy per cent identified as full-time or part-time professionals rather than hobbyists.

Within that population the findings are consistent. Twenty-one per cent had lost a job to a synthetic voice, up seven points in twelve months. Thirteen per cent had agreed to have a synthetic version of their voice made. Nine per cent had found a synthetic copy of their voice, or of a character they had voiced, used in a professional project without their knowledge.

The seven-point jump followed a collapse in the price and skill needed to make a synthetic voice, not an improvement in the voices themselves. Cloning a usable one now takes a short sample and a few minutes, on consumer products at consumer prices. Producers with no technical skill are doing what needed a specialist two years ago.

The same pattern is visible outside voiceover. A Harvard working paper reported in May 2026 that entry-level hiring at firms adopting generative AI has fallen by roughly 80 per cent per quarter since 2023, across all sectors, and that those firms are not moving beginners to harder work but removing the posts.

Where the displacement has landed

Segment Exposure What is happening
Telephony and IVR Very high Effectively conceded. Systems are generated and updated in software, and the cost of a re-record is now zero
High-volume e-learning and compliance modules High Bulk scripts, frequent revision, no named performer. The economics favour synthesis and the buyer rarely resists
Low-budget corporate and explainer video High The commodity end of the market, where price is the only criterion
Radio imaging, promos, station fill High and rising Being taken in-house as much as by AI; a US television group cut voice talent across more than 260 stations
Bulk localisation and dubbing Very high The hardest-hit segment worldwide, and the one with the fewest protections
Self-published non-fiction audiobooks High Audible's Virtual Voice took the long tail almost entirely
Branded and character audiobook work Low Buyers pay for performance, and listeners are moving away from AI narration, not towards it
Regulated medical and pharmaceutical narration Low Approval chains, liability and pronunciation risk make a briefable human cheaper than an error
Broadcast advertising with a named voice Low The voice is part of the brand asset and is contracted as such
Live-directed sessions of any kind Low The value is in the direction, which the machine cannot receive

Two patterns run through that table. Exposure tracks the absence of a person: wherever nobody needed to know whose voice it was, nobody now needs a voice. Exposure also tracks the absence of consequence. Where a mistake costs the client money or exposes them to a regulator, they buy a human who can be briefed, corrected and held to it.

Why the volume figures overstate the threat

The most quoted numbers in this debate count titles, not money, and the gap between the two is the single most useful finding in this research.

Audible's Virtual Voice, its self-service AI narration for Kindle Direct Publishing titles, passed 50,000 labelled titles during 2025, having stood at about 7,000 at the end of 2024. That is a flood by any measure. Yet in the Audio Publishers Association survey conducted by Edison Research and released on 5 June 2026, AI-narrated titles accounted for 0.03 per cent of audiobook sales revenue in 2025, in a US market worth 2.43 billion dollars and growing 9 per cent that year. Fifty thousand titles earning three-hundredths of one per cent of the money describes a long tail being filled in, mostly with books nobody would have paid a narrator to record.

Listener appetite is moving away from the technology as well. In the same survey, of 1,005 US adults who listen to fiction audiobooks, willingness to try an AI-narrated title fell from 70 per cent in 2025 to 61 per cent in 2026. Only 16 per cent had listened to one, and among those who had, preference for AI narration fell year on year.

Where the profession's defence is weaker than it claims

The industry's favourite argument is that listeners can always tell. They cannot, and it is worth Anita knowing this before a client tells her.

WPP Media, the sound branding agency amp, Choreograph and MediaScience ran a neuroscience study published on 27 October 2025 with 55 US participants aged 18 to 30 and 45 or over, in a simulated living room. Participants correctly identified generic AI voices in 42 per cent of individual sentences and 47 per cent of complete advertisements, which is close to guessing. On the measures advertisers buy, brand engagement and purchase intent, synthetic voiceovers matched human ones. Participants rated voices they believed to be human as more relatable and more emotional, whether or not the voice was human.

That is a small sample in one country, and it tests short advertising rather than long-form narration, where a synthetic voice has more chances to fail. Its direction is still clear enough to retire the detectability argument. The claim that survives is about provenance rather than about ears: buyers care where a voice came from, whether or not anyone can hear the difference.

What buyers say they want

Two buyer-side surveys point the same way, and both come from companies with a commercial interest in the answer, which should temper how much weight they carry.

Censuswide surveyed 700 business leaders and consumers for the Voices.com Amplified 2026 report. Seventy-nine per cent of those deciding on voice AI said inauthentic AI voices damage brand perception, and the same proportion said it matters that an AI voice comes from real, attributed talent. Seventy-seven per cent of business leaders valued exclusive voice licensing. Only 29 per cent of companies had deployed customer-facing voice AI at all, against 55 per cent of consumers using voice as their main way of interacting with AI systems.

The training body Gravy for the Brain reports a similar hesitancy among buyers. Roughly a quarter had experimented with AI voices, and most of those used it on fewer than a quarter of their assignments. About a third of buyers grew more open to AI over the year and about a third grew more hesitant. Human-likeness, rather than cost or speed, was the top consideration when they did use it.

Compulsory disclosure appears to cost brands less than the industry assumes. The Internet Advertising Bureau surveyed 505 US consumers aged 16 to 43 and 104 advertising executives between October 2025 and January 2026, and found 73 per cent of those consumers saying a clear AI disclosure would either raise their likelihood of buying or make no difference to it. The European Union has required that disclosure since 2 August 2026. Where a voice came from is therefore becoming visible in the market rather than only in the contract.

The law, and what it does for a British voice

Three jurisdictions are moving at three speeds.

The European Union is furthest ahead. Article 50 of the AI Act requires providers of systems generating synthetic audio to mark that output in a machine-readable form, and requires deployers to disclose synthetic content to a person at first exposure, in a way perceivable without special tools. Those obligations apply from 2 August 2026, with a transition to 2 December 2026 for systems already on the market before that date.

The United States protects performers mainly through contract. SAG-AFTRA's 2025 Interactive Media Agreement sets bargained minimums for digital replicas, pays vocal replica work by generated line of roughly ten words, and lets performers withdraw consent for new generation during a strike. The commercials contracts require consent for every use of a digital voice replica and let performers set their own price above the minimum. Federal statute is still pending: the NO FAKES Act and the NO AI FRAUD Act are before Congress, and several states have acted on their own.

The United Kingdom has the weakest position of the three. There is no standalone right in a person's voice. A claimant has to assemble a case from passing off, defamation, data protection and copyright, none of which was designed for this. On 28 August 2026 more than eighty performers wrote to the Prime Minister asking for a statutory right over the use of a person's voice, held by everyone rather than only by working performers. The signatories included Matt Lucas, Hugh Bonneville, Nicola Coughlan, Siobhán McSweeney and Luke Evans, and the campaign behind the letter, Save Our Voices Now, was co-founded by the audiobook narrator Alice Sockett and the performer Peter Caulfield. The government has so far offered a consultation rather than legislation. Equity held an indicative ballot in late 2025 in which 99.6 per cent of those responding said they would refuse digital scanning without contractual protection, and the union has pursued claims for more than twenty members whose voices were allegedly used without permission.

The practical consequence for Anita is plain. Until Parliament acts, her contract is her protection, and nothing else is. Every engagement she takes should say in writing whether the recording may be used to train or generate a synthetic voice, and the default answer should be no unless she is paid separately for it.

Medical work, considered on its own

Anita has named medical advertising and demos for organisations as her main target. That is a better-defended position than most, for structural reasons rather than sentimental ones.

Regulated promotional material passes through medical, legal and regulatory review. The FDA has given AI no separate compliance lane: the same rules on fair balance, claim support and risk presentation apply whether a human or a model produced the content, and the agency has been pressing on whether risk information is genuinely audible and comprehensible rather than merely present. In the United Kingdom, the ABPI Code and the Prescription Medicines Code of Practice Authority govern promotion to health professionals, and prescription-only medicines may not be advertised to the public at all. Most of the medical voice work available to a British artist is therefore facing health professionals, congresses, device demonstrations, patient information and internal training rather than consumer advertising.

In that setting a mispronounced drug name or a misplaced emphasis in a risk statement is a compliance problem rather than an aesthetic one, and correcting it sends the material back through review. A human who can be briefed on a pronunciation list, corrected in session and held responsible avoids that. Terminology density works the same way: medical scripts carry anatomical terms, drug names, acronyms and abbreviations that synthetic systems still mishandle without hand-tuning, and hand-tuning erodes the cost advantage that justified using them.

Healthcare marketing is meanwhile shifting away from polished scripted voiceover towards physician-recorded clips, patient testimony and care-team explainers, on the finding that unpolished material converts better on personal health topics. That reduces demand for narrated brand video whoever narrates it, and it has nothing to do with AI. Anita's exposure in medical work is more likely to come from that shift than from synthesis.

Audiobooks and Hindi

Audiobooks are the clearest example of a segment splitting rather than falling. The bottom has gone to Audible's Virtual Voice and to services charging tens of dollars for a full-length book against thousands for a human recording. What remains for humans is fiction with multiple characters, children's books, literary work, and any title where the publisher expects to earn back a production budget. Audible and ACX have also opened a beta letting selected narrators create licensed replicas of their own voices, keep approval rights and take a title-by-title royalty share, with the result labelled as AI on the product page.

Hindi cuts both ways, and Anita should think about how she frames it. Demand for Indian-language content is very large and growing. India is also where AI dubbing has been adopted hardest and where performers have the least protection: the Association of Voice Artists of India has said the jobs are being annihilated, no binding industry agreement exists, and studios are cutting production costs to a fraction of previous levels using synthetic dubbing at scale. Offering Hindi as cheap bulk localisation puts her in direct price competition with that. Offering English and Hindi from one medically literate performer, under a single UK contract with one person accountable for both, does not.

Licensing her own voice

There is a second revenue line, and the question it raises has to be settled before launch rather than after the first offer arrives.

Marketplaces now pay voice artists when their licensed clone is used. ElevenLabs reports having paid five million dollars in total to contributors to its voice library in under two years, which across the whole library is modest per head; realistic individual earnings run from negligible for a general voice to a few hundred dollars a month for a well-defined niche, with a small top tier earning more. Voices.com is building its business on consent-based, attributed voice licensing precisely because enterprise buyers are now asking where a model's voice came from.

The trap is irreversibility. Once a model of a voice exists, audio already generated from it cannot be recalled, and a licence with loose scope or no end date cannot be walked back by regret. If Anita licenses, she should license narrow: named uses, a stated term, a right to withdraw future generation, and an explicit exclusion of political, adult, medical-claim and endorsement content. If she does not want to license at all, the site should say so, because that itself is a selling point to buyers who care about provenance.

What follows for someone launching now

The threat to Anita is real but it is not the threat she will have been told about. She is not competing with AI for the work she wants. She is competing with AI for the work she does not want, and losing it, which does two things: it drives the going rate for that work below anything a person can live on, and it removes the rungs a beginner used to climb.

That has four consequences for how she launches.

  1. Enter above the floor. Do not price against synthesis on the commodity end, because that fight is lost and winning it pays badly. Price and pitch against the work where a human is the point.
  2. Make the specialism the whole proposition. Medical narration with acting training, correct clinical pronunciation and a Hindi second language is a defensible position. "Voiceover artist available for all projects" is not.
  3. Make provenance visible. State on the site that she performs every recording herself, state her policy on AI licensing and voice cloning, and state it in words a buyer searching for a human voice would type. Third-party human-voice certification marks now exist and are worth considering once the site is live.
  4. Replace the missing ladder deliberately. The cheap jobs that used to teach direction, studio discipline and turnaround are gone. Coaching, a properly treated recording space and directed practice have to substitute for them, and buyers can tell the difference within one audition.

Four of these points are now questions in Anita's website requirements. Three of them concern her voice: whether the site states a policy on AI and cloning, whether she will license a synthetic copy and on what terms, and how the demos should be served given that public audio is training material. The fourth asks which search terms a buyer looking for a human would type.

How far to trust these figures

Being self-selected and US-weighted, the NAVA survey measures the professional community that answers surveys rather than the whole trade. Voices.com and Gravy for the Brain both sell to voice artists or to the people who book them, so their framing follows their interest, and the WPP neuroscience study rests on 55 participants in one country. Treat market-size forecasts for AI voice as marketing rather than measurement: they run from roughly two billion dollars in 2026 to a hundred billion by 2035 depending on who is selling the report, and none of them is used above.

Two findings hold up better than the rest, because they have real denominators and nobody selling behind them. Twenty-one per cent of 1,379 working voice artists lost a job to a synthetic voice in a single year. AI narration took 0.03 per cent of US audiobook revenue in 2025 while occupying more than 50,000 titles.

Sources

Surveys and research:

Law and industry agreements:

Segments and platforms:

Enter your search text in the box above

Select a result to preview