The Vanished Hello: What Voice Cloning Costs Trust Between Strangers

The instruction arrived on 7 September 2026 in the Guardian's Pass notes column, filed under Social etiquette, sandwiched between a joke about Alexander Graham Bell's preferred greeting and a gag about being cloned into a French person. “If you're called by an unidentified number, the advice is to remain silent until you are sure that a friendly human being is on the other end of the line,” it read. “Not saying hello first could save you a whole heap of trouble.”

Read once, it is a light item about manners. Read twice, it is a security advisory dressed as a lifestyle note, and what it advises is the abandonment of the single most widely performed speech act in the history of telecommunications. The column was explicit about the danger. Scammers, it said, can now replicate your voice with artificial intelligence, cloning it from your telephone greeting and using it to beg your loved ones for money, and it credited coverage of the risk to the consumer watchdog Which?, the BBC and the Guardian itself. Asked what the best defence was, the column answered with a single word. Rudeness.

There is something quietly astonishing in that. The remedy is not a change to the network, not a certificate attached to a call, not a duty placed on the firm that sold the synthesis tool or the carrier that terminated the traffic. It is a change to you. You are asked to stop speaking first to people you do not recognise, and to hold that silence until you have satisfied yourself, from breath and hiss and hold music, that something human is on the other end. The burden is free to issue and expensive to carry, which is a large part of why it was issued.

The column's other culprit, generation Z, had already done some of the work. It cited a survey finding that 40 per cent of 18 to 24 year olds consider it acceptable to answer the phone with no greeting at all, and noted that the New York Times had reported young people substituting “Hi”, “What's up?” or “Yeah?” for the full formula. Caller ID made the ritual redundant for anyone who already knew who was calling. Voice cloning is finishing it off for everyone else, including the people for whom caller ID never resolved anything, because the number on the screen has been forgeable for two decades.

Three days later, researchers published the first serious count of how much of the inbound telephone stream is now a machine talking. Until that paper landed, the argument had run almost entirely on anecdote. What the counters found is both worse and stranger than the advice implies.

What Edison Actually Proposed in August 1877

Hello is not an old word doing a new job. It is a young word recruited for a specific industrial purpose, which it has performed for almost exactly 149 years.

On 15 August 1877, Thomas Edison wrote to T. B. A. David, president of the Central District and Printing Telegraph Company of Pittsburgh, and proposed hello as the proper way to answer a telephone. The Hagley Museum and Library, which holds a substantial part of the American industrial archive, records the date and the recipient. Edison's suggestion displaced the greeting recommended by Bell, who had invented the instrument the year before and who favoured “Ahoy, ahoy”. Bell, by most accounts, went on using ahoy for the rest of his life and lost.

The word itself was older than the telephone but not by much, and it was not a greeting. It descended from holla, hollo, halloo and hillo, noises used in English since at least the sixteenth century to grab the attention of someone at a distance: a shout across a field, a hail to a boat, a call to a dog. Its function was acoustic rather than social.

That is precisely why it worked on a telephone. The early instrument had no ringer worth the name and no way to signal that a line had opened. Hello was a test tone with a meaning attached. It established that the channel was live and that a person was at the other end of it. For the next century and a half those two jobs travelled together so reliably that nobody noticed they were separable.

They are separable now. That is the whole of the problem in one sentence. A live channel no longer implies a person, and a human voice no longer implies a human being, so the word that used to verify both at once verifies neither. Worse, it has been turned around. The test tone has become a sample. A word designed to prove presence has been recast as evidence handed to an unknown party.

This is not the first time telephone trust has had to be rebuilt. For decades the network's trust model was a human being: the operator, who knew the subscribers and could say who was on the line. Direct dialling removed her. Caller ID, which American carriers were required to offer from the late 1980s, was the replacement, an automated identity claim presented before the greeting. Voice over internet protocol then made that claim trivially editable, and by the mid 2000s consumer services were selling spoofed caller ID openly. Each layer has failed the same way. An identity claim that was cheap to make became cheap to fake, and the cost of the failure was pushed to the person holding the handset. Hello is simply the last layer, and the oldest.

Eleven Bait Numbers and the First Hard Count

On 10 September 2026, a team led by Xingyu Shen, with Simiao Ren as senior author and colleagues working under the affiliation Scam AI, part of Reality Inc, posted a paper to arXiv titled “The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls”. Its premise is a gap in the record. In February 2024 the Federal Communications Commission placed AI-generated voices under the Telephone Consumer Protection Act. More than two years later, the authors note, no peer-reviewed measurement said how much unwanted call traffic is placed by a machine, or how much of that machine speech is synthesised rather than replayed. The policy existed without a baseline.

Their instrument was an interactive voice honeypot: language-model personas answering real American telephone numbers, with the caller recorded on a separate track so the persona's own speech could not contaminate the analysis. The numbers were baited by submitting fabricated contact details into third-party lead-generation forms, injecting false records into the data brokerage funnels that feed outbound calling operations. Over 66 days the honeypot recorded 10,987 calls. Eleven days on which the stack answered silently were set aside, leaving 7,233 calls on which the persona actually said hello.

Three instruments read each opening. An audio fingerprint looked for the same recording appearing on other calls. A commercial synthetic-speech detector scored the caller's first ten seconds. Blinded listeners checked what it flagged.

The decomposition is the finding. Of the 7,233 greeted calls, 13.8 per cent opened with a recording the fingerprint also heard on another call. A further 13.1 per cent opened with fresh audio the detector labelled synthetic. Another 9.9 per cent opened with a caller who never spoke at all, staying on the line for a median of sixteen seconds in silence, which is what a predictive dialler produces when it has placed more calls than it has agents free. Some 54.2 per cent opened with fresh audio the detector labelled human, and 9 per cent could not be scored.

Add the two audio-evidenced machine rows and at least 26.9 per cent of answered calls opened with a machine voice, with a 95 per cent confidence interval of 24.3 to 29.8 per cent. Add the silent connections and a further tenth is almost certainly machine-placed. On a bait number in the middle of the American lead-generation economy, something between a quarter and two fifths of everything that answers your hello is not a person.

Two further results cut against intuition. Synthetic speech concentrated not in fraud but in ordinary commercial nuisance. Calls the study's classifier labelled spam, meaning unsolicited but non-fraudulent lead generation, were 33.8 per cent synthetic. Calls labelled scam were 21.1 per cent, falling to 15.9 per cent under a narrower criterion requiring a credential request. A rule argued as an anti-impersonation measure has landed, in volume, on telemarketing. The authors read the inversion as synthesis following volume economics: a fixed script benefits when the marginal cost of another call approaches zero, whereas extracting credentials from a resistant target still requires improvisation.

And almost nothing announces itself. Genuine self-disclosure of automation occurred on 8 of 1,816 detector-flagged calls, 0.44 per cent. Meanwhile 17 per cent of callers disclosed that the call was being recorded. Callers already read compliance preambles. The omission is specific to the speaker.

The Detector That Cannot Hold a Steady Line

The same paper is unusually candid about the limits of its own instrument, and those limits are the real story of synthetic speech detection in 2026.

Present the detector with the identical waveform on two different calls and it lands on opposite sides of its own threshold 13.6 per cent of the time. Eleven human listeners, auditioning what the detector flagged, confirmed 54.4 per cent of it. Agreement between those listeners was, in the authors' words, fair at best, with a mean pairwise coefficient of 0.267. The listening pool contained no known-human control audio, so a listener who answered “synthetic” to everything would have scored perfectly on the gold set. The false-negative rate was not estimable at all, which means the 54.2 per cent labelled human is a ceiling of unknown height. When the window was widened from ten seconds to twenty, both contested and control clips drifted towards “synthetic”, so even that precision figure is partly a property of the stimulus rather than of the speech.

This is not an artefact of one commercial product. AASIST, a standard academic anti-spoofing architecture, falls from an in-domain area under the curve of 1.00 to 0.43 on in-the-wild audio. The Deepfake-Eval-2024 benchmark, assembled by Nuria Alina Chandra and colleagues including Oren Etzioni from 56.5 hours of audio circulating on real platforms, found open-source audio detectors losing 48 per cent of their area under the curve relative to academic benchmarks, with the maximum across modalities reaching 0.58. Codec compression, which is what telephony does to every call, costs a mean 5.30 points of equal error rate.

The community's own flagship evaluation agrees. In ASVspoof 5, described by Xin Wang, Hector Delgado, Hemlata Tak and colleagues, baseline countermeasures scored equal error rates above 29 per cent against crowdsourced attacks including adversarial filtering. The top five submissions got below 15 per cent, roughly a halving, which is real progress and still nowhere near a figure you would accept from a lock.

Humans are no better. Kimberly Mai, Sergi Bray, Toby Davies and Lewis Griffin, writing in PLOS One in 2023, played genuine and synthetic audio to 529 people in English and Mandarin. Listeners identified the deepfakes 73 per cent of the time. Telling them what to listen for helped only slightly. The companion survey posted to arXiv in August 2026 by Chengzhe Sun, Tianle Yang and Siwei Lyu makes the structural point: synthetic voice detection is harder than image or video deepfake detection because phonetics, prosody and auditory perception are harder to model than pixels.

So the advice to stay silent is not arbitrary. It is what is left when detection does not work, the network will not authenticate, and somebody has to be told to do something.

The Thin Evidence Under the Silence Advice

Here is where honesty costs something. The mechanism the advice guards against, a criminal harvesting your hello and building a working clone from it, is the least well evidenced link in the chain.

The tooling does not obviously support it. ElevenLabs, whose instant voice cloning is the reference product in this market, recommends one to two minutes of good audio for an instant clone and 30 to 180 minutes for professional cloning, which trains a custom model. Its documentation notes that some users get excellent results from as little as thirty seconds. Thirty seconds is not hello. It is a voicemail greeting, a podcast appearance, a wedding speech on a public feed, or a voice note forwarded to the wrong group chat.

The precedent is also unflattering. In early 2017 American outlets reported a wave of “Can you hear me?” calls, said to record the victim's “yes” for use in authorising fraudulent charges. Snopes, checking with the Better Business Bureau, the Federal Trade Commission and the Consumer Federation of America, could find no record of anyone being defrauded that way. The scam was real as a call pattern and unproven as a theft mechanism, and the advice it generated outlived the evidence by years.

None of which makes the underlying threat imaginary. The industrial supply of cloning capability is real and lightly policed. Grace Gedye, a policy analyst at Consumer Reports, tested six voice cloning products in March 2025: Descript, ElevenLabs, Lovo, PlayHT, Resemble AI and Speechify. Four of the six asked researchers only to tick a box confirming they had the legal right to clone the voice. Four needed nothing beyond a name and an email address to open an account. Gedye's assessment was that there is a good argument some of what these companies offer runs afoul of existing consumer protection law.

So the accurate statement is narrower and more uncomfortable than the advice. Your voice is probably already available in sufficient quantity to clone, from sources you published yourself, and a stranger's telephone call is a poor way to acquire it. What answering an unknown number reliably does is confirm to an automated system that a human being holds the line, which is itself a saleable fact. The honeypot study shows what happens next: the researchers' first bait number, seeded in May, was receiving 59.4 per cent synthetic-voiced calls by July, while ten numbers seeded fresh in July ran at 18.6 per cent. Matched on how long each number had been in circulation rather than on the calendar, the gap vanished. Exposure is what drives prevalence. A number that has been answered gets called more.

Silence, then, is decent operational hygiene aimed at the list-building layer, sold to the public as a defence against voice theft. The gap matters, because a defence justified by the wrong mechanism cannot be evaluated, and cannot be retired when it stops working.

The Call That Came From the Sheriff's Own Number

On 9 September 2026, two days after the Guardian column, WEAR-TV reported the case that shows the limits of the advice most sharply. A woman in Crestview, in Okaloosa County, Florida, came close to losing nearly 25,000 dollars to a caller claiming she had missed jury duty.

The call did not present an unknown number. It appeared to come from the sheriff's department's main line. The caller identified himself using the name of a real deputy. Okaloosa County Sheriff Eric Aden described what his office is now up against in two sentences that deserve to be read as a technical summary. “They've gotten better and better to where they're using artificial intelligence using voice overrides,” he said. “They've gotten even better where they're spoofing legitimate numbers.”

She was instructed to deposit the money into what the caller described as a federally approved depository, which turned out to be a bitcoin machine inside a petrol station on Miracle Strip Parkway. She got that far. What stopped her was a printed disclaimer about jury duty scams on the machine itself. She drove to the sheriff's office instead, where the scheme was confirmed. Aden's own guidance was blunt about what the county never does: they do not ring to tell you that you have missed something, and they certainly never ask for a payment to make trouble go away.

Now apply the silence protocol to that call. It fails completely. The number was not unknown, it was the sheriff's. The name was not invented, it was a serving deputy's. Waiting for the caller to speak first would have produced exactly what the attacker wanted: an authoritative voice, identifying a real institution, from a number the recipient could verify on a website. The defence that saved 25,000 dollars was a sticker on a cash machine, placed there by somebody who had concluded the failure would happen downstream and had to be intercepted at the till.

That is the pattern worth holding on to. The honeypot data and the Florida case describe two different attacks that public discussion has merged. One is industrialised and mostly commercial: high volume, low value per call, synthetic voice used because it is cheap. The other is targeted and socially engineered, often conducted by human beings reading a script, with a cloned or merely confident voice as one prop among several and spoofed caller ID doing most of the identity work. Advice built for the first does nothing for the second, and the second is where the money goes.

UK Finance recorded 1.28 billion pounds lost to payment fraud in 2025, reported on 15 June 2026, with authorised push payment losses up 19 per cent to 576.4 million pounds across 248,070 cases, and 17 per cent of those cases originating through telecommunications channels. Ruth Ray, managing director of economic crime at UK Finance, put it without decoration: almost 1.3 billion pounds was stolen again last year, and it is clear the underlying problem is not being tackled effectively enough. In the United States, the FBI's Internet Crime Complaint Center logged 20.9 billion dollars in reported losses for 2025 across more than a million complaints, with 7.7 billion of that reported by people aged 60 and over, an increase of roughly 60 per cent on the previous year.

Why Nobody Asked the Network to Fix It

The FCC's declaratory ruling of 8 February 2024, adopted unanimously six days earlier as FCC 24-17, held that calls made with AI-generated voices are “artificial” within the meaning of the Telephone Consumer Protection Act. It took effect immediately and it was worth doing. It was also narrower than its headlines. As the honeypot authors point out, it created no new offence and banned nothing outright. It resolved a definitional question and attached existing machinery to synthetic speech: consent, identification of the responsible entity, opt-out rights, and enforcement through forfeitures, carrier blocking and state attorneys general.

Two omissions define its reach. It did not require a caller to say that a voice is a machine, which is why the study's 0.44 per cent disclosure rate is not a compliance failure but a baseline for a rule that does not exist. The Commission proposed that duty in August 2024, in a Notice of Proposed Rulemaking numbered FCC 24-84, and two years on it has not been adopted. And it did not change the unit that enforcement acts on. Blacklists, reputation scores, carrier analytics and consumer call blocking all identify the telephone number. The study found the number to be the perishable part: the typical number in a campaign places all of its calls within a single day and is never seen again, while the campaign itself, the same opening script dialled from a stream of fresh numbers, runs on for a further fortnight. One recorded compliance notice opened calls across six campaigns. One synthetic voice served nine. Blocking a number removes almost none of an operation's remaining life.

Enforcement shows the problem. Lingo Telecom, the carrier that transmitted the AI-generated Joe Biden robocalls into the New Hampshire primary in January 2024, settled with the FCC for one million dollars. The consultant who commissioned them, Steve Kramer, was hit with a six million dollar FCC forfeiture in September 2024, told the Associated Press he would not pay it, and in June 2025 was acquitted by a Belknap County jury on eleven felony voter suppression charges. The most politically salient synthetic voice call in American history produced one settlement from an intermediary and one unpaid fine.

The authentication route has fared no better on this side of the Atlantic. Ofcom consulted on calling line identification authentication, the British equivalent of the American STIR/SHAKEN framework, and in early 2024 declined to mandate it. Peter Farmer, writing for the operator Simwood in March 2024, summarised the reasoning: the costs of implementation, the complications of roaming and the intractability of calls originating abroad outweighed the benefits, since overseas operators are under no obligation to follow British verification rules. His own objection is the one that matters most here. No amount of cryptographic attestation addresses spoofing if there are no consequences for spoofers. Ofcom instead extended call blocking to international calls presenting British numbers, closing one loophole and leaving the architecture unauthenticated.

Meanwhile the honeypot's own traffic shows why authentication is not a side issue. The rise in synthetic calls to its numbers lived almost entirely in traffic presenting caller ID in the bait numbers' own 484 area code, so that each call looked local. The identity layer is not merely weak. It is actively used as an attack surface.

The People Who Cannot Afford to Screen Their Calls

Silence is not free. It is a luxury good, and the market has sorted who can buy it.

Hiya, which surveyed more than 12,000 consumers across six countries for its State of the Call report in 2026, found that 86 per cent of calls from unknown numbers now go unanswered, that one in three respondents report receiving deepfake calls, and that average losses to phone scams run above 800 dollars. Consumers in that survey expected mobile network operators to carry the financial responsibility for scam losses. The screened telephone is not a proposal. It is already the default behaviour of the overwhelming majority.

Now consider who cannot join in. Police Scotland's guidance on voice cloning scams advises people to avoid answering calls from unfamiliar numbers, to ask their provider to block anonymous calls, to ring back on a known number, to agree code words with family and to keep audio off public feeds. Every item is sensible. Every item assumes a life in which unexpected calls are threats rather than logistics, which is not most lives.

That assumption breaks immediately in British healthcare. Portland Road Practice, an English GP surgery, publishes a did-not-answer policy which asks patients to ensure their phone accepts calls from unknown numbers, because clinicians may ring with no caller ID or from a non-surgery line. Miss the call and you get one text and one further attempt. Miss both and you must rebook the slot yourself, and the missed call is recorded in your clinical record as a failed appointment. Persist and the practice will require future appointments to be face to face. Derbyshire Healthcare NHS Foundation Trust moved off withheld numbers in February 2023 precisely because, as it explained, some patients' phones block withheld numbers to stop cold calling, and many people do not feel comfortable answering calls when they do not know the number.

So a patient on a waiting list is instructed by one arm of the state to answer unknown numbers and by another to avoid them. The conflict is resolved nowhere. It is devolved to the individual, along with the consequences of guessing wrong.

The same devolution lands on anyone whose income arrives from strangers. A self-employed electrician, a supply teacher, a locum, a courier, a carer covering shifts: for these people an unknown number is a job, and a policy of silence is a policy of lost work. It lands on people in insecure housing waiting for a landlord, a council officer or a hostel place, and on people inside immigration processes, where an unreturned call can carry procedural weight. It lands hardest on the isolated, for whom an unexpected call is contact rather than threat, and for whom answering in wary silence rewrites the one remaining pleasure of a ringing telephone.

The advice is free to give. It is expensive to follow, and the bill is itemised by class, health and employment status.

A Biometric That Nobody Volunteered

Underneath all of this sits a category error that the industry made a decade and more ago and is only now unwinding.

The voice became an authentication factor without anybody consenting to make it one. HSBC launched Voice ID for telephone banking and had roughly half a million customers enrolled by 2017, when the BBC's Click programme tested it. The reporter Dan Simmons enrolled, and his non-identical twin Joe then talked his way into the account, failing seven times before being admitted on the eighth. A Click researcher found the system permitting more than twenty attempts across twelve minutes. HSBC said it would review the system's sensitivity and noted that twins share a similar voiceprint.

That was a decade before consumer-grade synthesis became a subscription product. What has happened since is demotion rather than abandonment. A voiceprint that can be approximated by a sibling was always a weak factor. A voiceprint that can be approximated by a subscription service is not a factor at all, and the detection literature explains why nothing can currently be bolted on to rescue it. If the best systems in an open evaluation still miss more than one attack in seven, a voice cannot carry an account on its own.

The epistemics should worry people more than the banking. A biometric is supposed to be a property you cannot help having and cannot easily surrender. Voice fails that test in an unusual direction. It is not a poor identifier. It is a broadcast identifier. You emit it constantly, in public, as the price of ordinary life, and every emission is a sample. Fingerprints require contact. Irises require proximity. Voice requires only that you speak within range of a microphone, and the entire built environment is now a microphone.

The Federal Trade Commission understood the shape of the problem early and tried to buy a way out. Its Voice Cloning Challenge, announced in late 2023, named four winners in April 2024: OmniSpeech's AI Detect, DeFake from Ning Zhang, which adds distortions to voice samples to frustrate cloning, OriginStory, which authenticates a voice as human by measuring biosignals in the throat and mouth at the point of capture, and a recognition award for Pindrop Security's real-time clone detection. Two of the four abandon detection entirely, authenticating the human at the source or poisoning the sample before it can be taken. The prize was an admission that catching a fake after the fact is the harder bet.

Who Profits From a Telephone Nobody Answers

There are places where this could be interdicted, and they are all upstream of the person holding the phone.

Authentication at the network layer is the obvious one, and the honest assessment is that it is necessary and insufficient. Cryptographic attestation of calling numbers would end casual spoofing of a sheriff's main line, the mechanism that nearly cost a woman in Crestview 25,000 dollars. It would not touch a lawfully originated call carrying a synthetic voice, and Ofcom's cost analysis was not obviously wrong. Provenance and watermarking sit in the same position. They mark the output of cooperating tools and are worthless against a model that has been fine-tuned, run locally or stripped of its markings by the codec on the way through the network.

Out-of-band verification is the most robust thing available and it is deliberately low technology. A callback on a number you already hold. A family phrase agreed in advance. In Britain, the 159 short code run by Stop Scams UK, which routes a caller directly to their own bank and has passed a million calls with coverage of more than 99 per cent of current accounts. None of these authenticate a voice. All of them stop treating the voice as the thing that authenticates, which is the actual fix.

Liability is the lever nobody wants to pull. British banks now reimburse most authorised push payment losses, and payment fraud still rose. The telecommunications channel carried 17 per cent of those cases and the telecommunications industry carried none of the cost. Voice synthesis firms that accept a ticked box as proof of consent carry none either. The finding that campaigns outlive their numbers by a fortnight means per-number enforcement is aimed, with real precision, at the cheapest component of the attacker's stack.

So who benefits from the current arrangement? Carriers, who continue to terminate and be paid for traffic that a quarter of the time is a machine reading a script, while the regulator declines to mandate the authentication that would make them prove otherwise. The lead-generation industry, whose synthetic calls are the most automated segment of the entire stream and which almost never discloses automation, because no rule requires it to. The synthesis vendors, who sell cloning for the price of a streaming subscription and an unverified attestation. Detection vendors, who sell to banks and enterprises rather than households, and whose products, by the measurement of the only field study we have, flip on identical audio more than one time in eight.

And the cost? It sits with a woman in Florida reading a warning label on a bitcoin machine. With a patient whose missed call has been logged as a failed appointment. With a tradesman who cannot afford not to answer. With everybody now expected to conduct the opening seconds of every call with a stranger in defensive silence, listening for the tell, performing an authentication task that no detector in the published literature can perform reliably, and absorbing the loss personally when they get it wrong.

Hello lasted 149 years. It was retired not by a vote, or a rule, or a standard, but by a lifestyle column recommending rudeness, because rudeness was the only remedy anybody was willing to make free.

Sources and References

1. The Guardian, “The death of 'hello': how gen Z and scammers put paid to the greeting”, Pass notes, Social etiquette, 7 September 2026. https://www.theguardian.com/lifeandstyle/2026/sep/07/death-of-hello-gen-z-a= nd-scammers
2. Xingyu Shen, Tommy Duong, Muduo Xu, Xiaodong An, Jiaqi Gan, Haoyuan Tang, Jamey Z. Liang, Siyu Zhang, Yan Zhang and Simiao Ren, “The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls”, arXiv:2609.11137v1, 10 September 2026, full text including Appendix A, Threats to validity, and Appendix B, Ethics. https://arxiv.org/html/2609.11137v1
3. Chengzhe Sun, Tianle Yang and Siwei Lyu, “A survey of AI-generated voices and their detection”, arXiv:2608.15411, 15 August 2026. https://arxiv.org/abs/2608.15411
4. Hagley Museum and Library, “On this date in 1877, inventor Thomas Edison suggested 'Hello' to the president of the Central District and Printing Telegraph Company of Pittsburgh”, Hagley Vault. https://www.hagley.org/research/news/hagley-vault/date-1877-inventor-thomas= -edison-suggested-hello-president
5. Federal Communications Commission, “In the Matter of Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts: Declaratory Ruling”, FCC 24-17, CG Docket No. 23-362, adopted 2 February 2024, released 8 February 2024. https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf
6. Federal Communications Commission, “In the Matter of Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts: Notice of Proposed Rulemaking and Notice of Inquiry”, FCC 24-84, CG Docket No. 23-362, adopted 7 August 2024, released 8 August 2024. https://docs.fcc.gov/public/attachments/FCC-24-84A1.pdf
7. CyberScoop, “FCC hits operative behind New Hampshire robocall with $6 million fine”, 26 September 2024. https://cyberscoop.com/fcc-fine-joe-biden-deepfake-new-hampshire-robocall-s= teve-kramer/
8. Associated Press, “NH jury acquits consultant behind AI robocalls mimicking Biden on all charges”, published by NBC Boston, 13 June
9. https://www.nbcboston.com/news/politics/nh-jury-acquits-consultant-be= hind-ai-robocalls-mimicking-biden-on-all-charges/3741539/
10. Peter Farmer, “On STIR/SHAKEN, Ofcom nailed it. On spoofing....?”, Simwood, 11 March 2024. https://simwood.com/2024/03/on-stir-shaken-ofcom-nailed-it-on-spoofing/
11. WEAR-TV, “Crestview woman nearly loses $25K in jury duty AI voice scam”, 9 September 2026. https://weartv.com/news/local/fort-walton-beach-woman-nearly-loses-25k-in-j= ury-duty-ai-voice-scam
12. Hiya, “State of the Call 2026”, survey of more than 12,000 consumers across six countries. https://www.hiya.com/state-of-the-call
13. UK Finance, “Fraud remains a national security threat as criminals steal almost =C2=A31.3 billion”, Annual Fraud Report 2026 press release, 15 June 2026. https://www.ukfinance.org.uk/news-and-insight/press-release/frau= d-report-2026-press-release
14. Jessica McMaster, “FBI Report: Internet Crime Losses Hit $20.9 Billion”, AARP, 16 April 2026, reporting the FBI Internet Crime Complaint Center's 2025 Internet Crime Report. https://www.aarp.org/money/scams-fraud/fbi-ftc-report-2025-losses/
15. Stop Scams UK, “159 passes one million calls”. https://stopscamsuk.org.uk/159-one-million-calls-stop-scams-uk-virgin-money= /
16. Police Scotland, “Voice cloning scams”, advice on scams and frauds. https://www.scotland.police.uk/advice/scams-and-frauds/voice-clonin= g-scams/
17. Thomas Claburn, “Consumer Reports calls out slapdash AI voice-cloning safeguards”, The Register, 10 March 2025, reporting research by Grace Gedye of Consumer Reports. https://www.theregister.com/2025/03/10/ai_voice_cloning_safeguards/
18. ElevenLabs, “Voice cloning” product guide, audio requirements for Instant Voice Cloning and Professional Voice Cloning. https://elevenlabs.io/docs/product-guides/voices/voice-cloning
19. Kimberly T. Mai, Sergi Bray, Toby Davies and Lewis D. Griffin, “Warning: Humans cannot reliably detect speech deepfakes”, PLOS One, 18(8):e0285333, 2 August 2023. https://journals.plos.org/plosone/article?id=3D10.1371/journal.pone.0285333
20. Xin Wang, H=C3=A9ctor Delgado, Hemlata Tak, Jee-weon Jung, Junichi Yamagishi, Nicholas Evans and colleagues, “ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale”, arXiv:2408.08739, 16 August 2024. https://arxiv.org/abs/2408.08739
21. Nuria Alina Chandra, Ryan Murtfeldt, Oren Etzioni and colleagues, “Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024”, arXiv:2503.02857, 4 March 2025. https://arxiv.org/abs/2503.02857
22. Snopes, “'Can You Hear Me?' Scam Warning”, fact check, 2017. https://www.snopes.com/fact-check/can-you-hear-me-scam/
23. Biometric Update, “BBC reporter and his twin dupe HSBC voice ID recognition system”, May 2017. https://www.biometricupdate.com/201705/bbc-reporter-and-his-twin-dupe-hsbc-= voice-id-recognition-system
24. Derbyshire Healthcare NHS Foundation Trust, “New number when we call you, to replace any withheld numbers”, 4 January 2023. https://www.derbyshirehealthcareft.nhs.uk/about-us/latest-news/new-number-w= hen-we-call-you-replace-any-withheld-numbers
25. Portland Road Practice, “Did Not Answer Policy”, NHS general practice patient guidance. https://www.portlandroadsurgery.nhs.uk/did-not-answer-policy
26. Federal Trade Commission, “FTC Announces Winners of Voice Cloning Challenge”, press release, 8 April 2024. https://www.ftc.gov/news-events/news/press-releases/2024/04/ftc-announces-w= inners-voice-cloning-challenge

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...