Artificial intelligence could help make legal services in Bangladesh more accessible. But before we entrust it with citizens’ legal problems, we must decide who is accountable when it gets them wrong.
Recently, the Legal Tech Hackathon 2026, supported by UNDP Bangladesh and its partners, brought university students together to develop solutions for more accessible, inclusive, and effective legal aid. Its AI for Legal Aid Services track included proposals for Bangla chatbots, voice assistants, legal triage, and document summarisation.
These tools could help people overcome barriers of cost, distance, information, and language. They could explain procedures, locate services, and help prepare documents. Yet moving from a promising prototype to a public service requires more than technical ingenuity.
Imagine a rural citizen who cannot afford a lawyer and consults a Bangla AI assistant about a land dispute. The answer is confident but wrong: the system misunderstands the facts, relies on outdated legislation, or invents a legal citation. The citizen acts on it and suffers legal or financial harm.
Who is responsible: the citizen, developer, platform operator, government agency, or reviewing lawyer? Unless that responsibility is clear, accountability has not been built into the system.
A strong benchmark score cannot answer this question. Justice concerns individual rights and consequences. A system that performs well in testing can still fail someone when accuracy matters most.
Not all legal AI carries the same risk. Finding a legal-aid office is different from deciding who qualifies for assistance. Explaining public legal information is different from offering personalised advice.
Bangladesh needs a risk-based approach. Lower-risk uses could include locating services, arranging appointments, and explaining publicly available procedures. These still require safeguards. Drafting assistance, legal research, case-information classification, and triage could fall into a medium-risk category, requiring stronger validation and meaningful human review. High-risk uses could include systems that influence legal-aid eligibility, prioritise legal problems, assess individuals in criminal or judicial contexts, or substantially shape judicial or quasi-judicial decisions. These need the strongest controls.
The dividing line should be the system’s potential impact, not its label. The greater the consequences for a person’s rights, the stronger the oversight must be.
International experience offers useful principles. The Council of Europe’s European Commission for the Efficiency of Justice adopted its Ethical Charter on AI in judicial systems in 2018, emphasising fundamental rights, non-discrimination, quality and security, transparency and fairness, and user control. The European Union’s AI Act treats certain AI systems used in the administration of justice as high-risk.
Singapore shows how innovation and governance can develop together. Its courts’ 2024 guidance requires users to check AI-generated material, verify legal authorities, and avoid fabricating or manipulating evidence. Experiments with AI-assisted case summarisation in its Small Claims Tribunals have avoided presenting the tool as a substitute for case-specific legal advice. In March 2026, its Ministry of Law also issued guidance stressing professional ethics, confidentiality, and transparency.
Bangladesh need not copy these approaches. It should recognise their central lesson: using AI does not remove human responsibility.
Legal AI may encounter details of domestic disputes, criminal allegations, property conflicts, and financial hardship. Citizens should know where their information is stored, who can access it, how long it is retained, and whether it is encrypted. They should also know whether it can be transferred abroad, shared with other institutions or used to train models, and what happens when they request deletion.
These questions must be answered before deployment, not after a breach. Access to justice should not require surrendering privacy.
Language presents another challenge. Fluent Bangla is not necessarily reliable legal Bangla. A system may sound convincing while misunderstanding legal terminology, regional expressions, or the context of a dispute. Small differences in wording can change legal meaning.
Testing must therefore be specific to Bangladesh. Evaluators should check whether cited laws exist and remain current, whether interpretations are sound and whether the system distinguishes law from opinion. They must also assess how it communicates uncertainty, handles high-risk situations, and refers users to qualified lawyers.
“Human in the loop” is not a safeguard unless the human has the competence, authority, time, and evidence to intervene. Who reviews the output? Are they a qualified lawyer? Can they inspect its sources? Are they responsible for approving it? Can the citizen challenge the result?
Someone clicking “approve” on hundreds of recommendations each day does not provide meaningful oversight. Human review must be a substantive check, not a procedural formality.
Bangladesh should establish a governance framework before scaling AI in legal services. It should define risk categories and accountability, require human review for high-impact uses, and set rules for collecting, storing, and reusing legal data.
Independent testing should cover accuracy, bias, security, and failure modes. Citizens should be told when they are interacting with AI, have access to qualified human assistance, and be able to challenge AI-assisted assessments. Serious failures should be reported, and systems should maintain audit trails. Procurement contracts should specify obligations for security, accountability, and auditability.
The hackathon can help put these principles into practice. Each prototype should explain what it does, the highest-risk decision it can influence, what data it collects, and which authoritative legal sources it uses. Teams should show how those sources are updated, how accuracy is independently assessed, and when users must be referred to a lawyer.
They should also answer the essential question: who is accountable when something goes wrong?
A prototype may work beautifully on stage. A public justice service must work when facts are incomplete, laws have changed, databases are unavailable, or a citizen disputes the outcome. It must protect users when the technology fails, not merely impress them when it succeeds.
AI could reduce information barriers and support overburdened legal professionals. Its purpose, however, should be to make justice more accessible without sacrificing fairness, privacy, or human dignity.
Bangladesh does not need to choose between innovation and governance. It needs governed innovation.
The question is not simply whether we can build an AI legal assistant. It is whether citizens can safely trust it, and whether they will be protected when it gets something wrong.
Nafiul Ahmad Rafi is the founder and director of Atlas AI Institute, a Dhaka-based think tank.
.......................................................................................................................................................................................
Views expressed in this article are of the author’s own and may not reflect the editorial stance of The Daily Star.
