Since the beginning of this month, much of the international press have carried out the same story: Jacob Coxon, a 27-year-old British researcher who spent three years on pretraining work at OpenAI and Anthropic, resigned on September 9 to sound a public alarm. He left two months before his Anthropic equity vested, forfeiting a massive corporate fortune to speak out.
Sitting opposite Anderson Cooper on CNN, Coxon revealed that the people building these frontier systems privately believe the technology could kill us all by the end of the decade, yet they are racing towards self-improving superintelligence regardless. Evan Hubinger, still inside the company, similarly estimates this existential risk above one in ten within the decade.
Crucially, Coxon made one point that most mainstream coverage missed: Today's systems are not an extinction threat. What frightens him is the sheer velocity at which they are beginning to improve themselves.
Warnings from these insider engineers deserve serious attention. But while Washington, Brussels, and San Francisco argue about a hypothetical intelligence explosion, Bangladesh has already experienced a much more real and immediate threat.
We do not need a sentient machine to damage our society. Everyday consumer software is already doing the work, and almost no one has noticed.
The bad actor from Bangladesh
Anthropic’s report, published on September 10, describes a lone operator working from a laptop in Gaibandha. Over roughly 16 months, he ran 29 rotated Claude accounts through a custom Python script and produced, by the company's count, at least 1,500 Bengali headlines, 300 fabricated narratives, and 1,500 image prompts.
The content was designed to boost one specific political faction while systematically undercutting its rivals. Crucial context disappeared amid the immediate commentary. The company reports no evidence that any political group directed or funded the work. It also states that it could not establish how widely the material circulated.
These unverified findings come solely from one private firm's internal logs. What remains beyond question is the changing equation of deception, and that is the shift deserving our attention.
A propaganda line that once required a coordinated infrastructure of writers, editors, and designers can now be run by a single individual with a basic internet connection.
The cost of producing a plausible, grammatically-clean falsehood in Bangla has fallen close to zero. When fiction becomes free to manufacture while verification remains expensive, the public square tilts. It does not tilt back on its own.
For years, the standard response has been to ask whether we can detect the fake. That approach fails on simple mathematics. A producer can generate a fresh variant faster than any classifier or fact-checker can categorize the last one.
The local defensive line in that race remains dangerously thin. Digitally Right's election study counted roughly 40-50 professional fact-checkers for a country of more than 170 million people.
Most newsrooms lack a dedicated verification desk. The threat is not that individual citizens fall for fakes, but that synthetic saturation ruins collective trust in the public square.
When any image or audio can be flawlessly synthesized, the authority of genuine evidence collapses. A study in the American Political Science Review found that political deepfakes are not inherently more persuasive than ordinary text rumours, but politicians who falsely cry deepfake successfully protect their standing.
A 2026 experiment by researchers at Würzburg, published in Psychology of Popular Media, explained why this tactic works. The false claim dramatically raises the chance that viewers misidentify genuine footage as fabricated.
Legal scholars Bobby Chesney and Danielle Citron named this the liar's dividend. The terminal damage of generative AI is not that citizens believe lies, but that they lose the will to believe the truth.
That erosion is especially dangerous in Bangladesh, where the information environment is fragmented and crowded with anonymous pages and unverified forwards.
Around the February 2026 election, fact-checkers logged a surge of AI-generated material. Research by Digitally Right documented fabricated videos and synthetic images deployed against women candidates and minority communities.
In response, the Election Commission barred using AI tools to distort facts, mislead voters, or compromise electoral integrity.
The state's primary answer has been reactive removal. Following the passage of the Cyber Security Act in April 2026, the government rapidly introduced draft amendments to force platforms to take down harmful AI-generated content within a fixed window.
The framework empowers the Bangladesh Telecommunication Regulatory Commission to block material and demand user data. The latest draft proposals reintroduce severe penalties, floating up to ten-year sentences for spreading rumours and disinformation online.
Rights groups, including Transparency International Bangladesh, have rightly warned that these sweeping changes risk a return to the repressive Digital Security Act era.
We now have a well-developed answer to one question: How do we take material down? Yet, we have almost no answer to another: How do we keep people believing material that is true?
The Draft National AI Policy 2026–2030 is not blind to this dilemma, explicitly emphasizing transparency, human oversight, and public trust.
Similarly, the Personal Data Protection Act covers the governance of data. But no statute can manufacture belief. A government can order a platform to remove a video, but it cannot order a citizen to trust the next one.
To heal this rift in trust, Bangladesh does not need faster censorship pipelines or wider legal nets. It needs a foolproof way to prove the creation of public-interest truth.
Rather than building a bloated, general-purpose AI regulator, the state should forge a tight, targeted national evidence chain for critical media.
For critical updates, like election tallies, official decrees, disaster warnings, public health guidelines, and state archives, institutions must publish digitally-signed media carrying tamper-proof source data straight from the point of origin.
News organizations should be supported, and for high-impact investigations eventually required, to preserve and display the original credentials of their footage. Platforms operating here should have to read those credentials and show users whether a circulating clip matches the signed original.
There is a real obstacle to admit: Most platforms re-encode uploads in ways that break the signature, so a duty to read credentials is worthless without a duty to preserve them. That is a negotiation with Meta, TikTok, or Google, not just a line printed in a state gazette. It should be the regulator's first demand.
Attribution, however, is no synonym for truth. An unedited video can still capture a politician telling a lie. It does not validate a narrative; it simply fixes the baseline.
It guarantees three basic facts:
- That the file is real;
- That the source is clear;
- That the content hasn't been altered.
That shifts the burden away from exhausting forensic guesswork, giving citizens, journalists, and courts solid ground to stand on.
None of this requires reinventing the wheel. The Coalition for Content Provenance and Authenticity publishes an open standard, now in its fourth major revision, with Adobe, Microsoft, Google, OpenAI, Meta, the BBC, and the Associated Press, among thousands of members.
American cyber authorities recommended content credentials to government agencies in a 2025 advisory; Google Search surfaces the data, OpenAI attaches credentials to image outputs, and Leica, Nikon, Sony, and Canon ship cameras that sign at the shutter. Adoption is real but patchy. Bangladesh's task is integration, not invention.
The individuals most vulnerable to synthetic information are not the uneducated but the time-poor. A garment worker checking her phone after a long shift, a farmer reading a flash-flood alert, a first-time voter trying to understand a candidate’s speech.
None of them can run forensics on every file that reaches them. A democracy cannot expect its citizens to be part-time investigators.
Nor can we accept an ecosystem where any inconvenient piece of journalism is waved away as an AI-generated fake. The first danger is mass deception. The second is blanket cynicism, and it is harder to undo.
Leaving this suspicion unchecked also creates a profound sovereignty problem for governance. Frontier AI firms have become the de facto border police of global digital behaviour, detecting and disabling accounts they judge malicious.
Who decides what counts as a suspicious pattern when a Bangladeshi researcher or student queries a foreign-hosted model?
Digital sovereignty cannot be outsourced to a corporate board in California. Concurrently, it cannot mean that our domestic security apparatus inspects the private digital interactions of every citizen.
The only viable standard, which the draft policy introduces, is that any algorithmic decision materially affecting a Bangladeshi must be human-reviewable, fully auditable, and open to appeal.
This framework must govern the foreign enterprise services that power our economy, not merely public sector infrastructure.
It is deeply contradictory to obsess over existential risks, like whether a future machine will declare humanity obsolete, while failing to secure the foundational evidence by which society determines what is true in the present.
The initial casualty of the AI era will not be physical; it will be the consensus that evidence is final. Once that epistemic baseline is lost, it cannot be salvaged by a superior language model, an expedited fact-check, or a restrictive censorship order.
A democracy can endure intense ideological friction over its future. It cannot endure a total fracture regarding its past. Bangladesh must establish this mechanism today, before reality itself is reduced to a permanent question mark.
Dr Sabbir Ahmad, a tech executive with global experience in digital connectivity, sustainable infrastructure and energy, now shaping Bangladesh's semiconductor landscape as the CEO of Silicon Array Ltd. Email: sabbir@ieee.org.