Sciences & Technology
Developing countries are writing AI laws they cannot enforce
Bangladesh's AI law was written by the companies it's meant to regulate
The steering committee for Bangladesh’s national AI policy lists three chief executives of the AI industry. When they help decide the rules that govern them, regulation becomes mere theatre
Published 22 September 2026
Bangladesh's National AI Policy 2026-2030 was drafted with guidance from a 29-member Steering Committee, with the full list published as an appendix to the policy itself.
Alongside government officials, university professors and a UN consultant sit Shahir Chowdhury, Raisul Kabir and Rahat Ahmed – three chief executives of major software companies and venture funds that the policy is meant to claim jurisdiction over.
There's Kabir's Brain Station 23, a major software and IT services exporter; Ahmed's Anchorless Bangladesh, an early-stage venture fund; and Chowdhury's Shikho, Bangladesh's best-funded education technology company.
Together, they make a meaningful share of the commercial AI activity operating in Bangladesh today.
Shikho's new AI chatbot in particular – given its scale and impact on children – is exactly the kind of system the National AI Policy was written to govern.
It is a product that would be considered high-risk, triggering mandatory impact assessments and oversight from the committee.
And yet, Chowdhury sits on the committee that decided what regulating that activity would actually mean.
We do not know and I am not suggesting that Kamir, Ahmed or Chowdhury asked for anything, but their presence raises some reasonable concerns about governance.
Their participation sits in the policy's own membership list, next to the committee's disclaimer that members served “on a voluntary basis” and that “no artificial intelligence tools were used in formulating the policy".
The second claim is almost certainly true. Given who was in the room, it is also beside the point.
The policy did not need an algorithm to produce an outcome favourable to the people who wrote it. It only needed those people to be present.
The theatre of AI regulation
Researchers who study AI regulatory capture – where dominant industries strongly influence the agencies built to police them – have a term for this.
It's called multi-stakeholder theatre.
Multi-stakeholder theatre is where a regulated company is invited to the table as a neutral expert under the guise of broad consultation, while its commercial stake in the outcome goes unstated and unmanaged.
Bangladesh's Steering Committee is close to a textbook case.
Its legitimacy rests on procedure: officials, academics and industry all had a seat. But procedure alone cannot tell you whether the industry seats were held by people with something to gain.
This pattern has been studied mostly in rich democracies, where a functioning press, competing law firms and well-funded advocacy groups can push back on an industry-friendly rule before it hardens into law.
Bangladesh has little of that ballast.
Politics & Society
Pausing AI gives democracy a chance to catch up
The country has few independent technologists outside the companies an AI policy would regulate, no dedicated AI policy press and a civil society sector already working under real pressure.
When a government needs people who understand machine learning well enough to help write law about it, it has almost nowhere else to look but the industry itself.
That is not a conspiracy. It is a resource constraint operating like a conspiracy, which may be even worse.
A conspiracy can be uncovered and punished, while a structural default just quietly repeats itself, policy after policy, unless someone decides to name it.
Whose law is it anyway?
None of this proves that Shikho, Brain Station 23 or Anchorless Bangladesh asked for anything specific in return, and I am not alleging they did.
The issue is structural, not personal. No one has to ask for anything directly.
When the people setting the rules also run the companies the rules cover, every ambiguous clause and every generous compliance deadline tends to resolve in one direction by default.
That is what AI regulatory capture looks like when it's working well.
It doesn't require a bribe. It only requires that the people in the room who understand the technology best are also the people who profit from how gently it ends up governed.
In the wake of recent calls for an AI slowdown, the lessons in Bangladesh are a warning for the US and the world – we need to be vigilant about who is in the room when these regulation talks occur.
In Bangladesh, the policy’s Independent Oversight Committee carries quasi-judicial powers, meaning it can issue binding findings about who has broken the law, not merely recommend that someone else decide.
Bangladesh does not have a clean record with instruments like this.
Amnesty International has documented how the country's cyber-security legislation, sold to the public as a safety measure, was used instead to prosecute critics and suppress dissent.
A quasi-judicial AI oversight body should be met with the same scrutiny.
It's about whose interests the law defaults to protecting once the cameras are gone. On the evidence we have, that default has rarely favoured the public.
Bangladesh needed this policy. The harms of ungoverned AI – in classrooms, in credit decisions, in facial recognition deployed with no legal basis at all – are real and are not waiting politely for the country to build perfect institutions first.
But a law does not earn public trust from the ambition of its language.
It earns that trust from who was kept out of the room while it was being written, and on that measure, Bangladesh's National AI Policy failed before its first clause was drafted.