Recently, some leading figures in the artificial intelligence (AI) industry, including Anthropic CEO Dario Amodei, have warned that AI capabilities are advancing faster than our ability to make them safe, reliable, and controllable.
This raises three questions:
- What has changed about AI?
- What risks does it create?
- How should Bangladesh respond?
For years, the same leaders pushed businesses to adopt AI quickly, and it worked. AI has become a mainstream business tool across industries, promoted as a major source of productivity and economic growth.
Until recently, most AI tools functioned as assistants. A user gave the system an instruction, the AI produced an answer, and a human decided what to do with it. The human remained in control.
That is changing. Newer AI systems are becoming more autonomous. Instead of answering a single question, an AI “agent” can be given a broader goal and determine the steps needed to achieve it -- searching for information, using software, writing and running code, and carrying out actions with less human involvement.
This shift, from AI as a tool to AI as an agent, is behind the concern.
The more autonomy a system has, the greater the consequences if it misunderstands an instruction, makes an error, is manipulated, or acts in a way its operator did not intend.
The concern is not that today's AI will suddenly become an independent decision-maker with its own agenda. It is that increasingly capable systems could eventually be given more authority than humans can effectively supervise.
Some researchers also worry about “recursive self-improvement,” in which AI systems help develop more capable systems, potentially advancing faster than researchers can test or control them.
There are also immediate risks. Powerful AI can be misused for cyberattacks and fraud. It does not need malicious intentions to cause harm; it only needs to be a powerful tool in the wrong hands or connected to an important system without adequate safeguards.
Autonomous agents raise the stakes because they can take multiple actions without a person approving each one.
The message from AI leaders is not “stop using AI.” It is that as these systems become more powerful, safety measures, human oversight, and governance must advance with them.
Why AI gets things wrong
AI is powerful, but it is not fool proof. Its output depends on the quality of its data, system design, and the context it receives. Human judgement remains essential.
If the data are inaccurate, outdated or poorly organized, AI can produce an unreliable answer with great confidence. It may also reproduce biases contained in historical data.
Generative AI has another key weakness. It can produce information that sounds plausible but is false. It may generate an incorrect statistic, invent a source, or provide a misleading explanation. This phenomenon, commonly called an AI “hallucination,” can be dangerous when such an answer is accepted without verification.
AI can also fail because it lacks important context. A sales forecasting system may recommend increasing production after analyzing years of sales data. But if a major customer cancelled an order yesterday and that information was not included, the forecast could be seriously wrong.
The lesson is simple: AI output should never be treated as automatically correct, especially when it informs financial, legal, safety, or employment decisions.
Matters for Bangladesh
Bangladeshi businesses cannot afford to adopt expensive technology because it is trendy. Many companies still struggle with basic data management. Records may be scattered across departments, stored in inconsistent formats, or divided between Bangla and English.
AI cannot fix these weaknesses. Poor data fed into a sophisticated AI system can and will simply produce mistakes faster and with greater confidence. Before investing heavily in AI, businesses should improve the quality, accessibility, security, and governance of their data.
A two-track response
Bangladesh does not need to panic about AI, nor should it rush blindly into it. The right approach is responsible acceleration: Move quickly where AI can improve productivity and competitiveness, but proceed more carefully where it affects people's money, safety, privacy, or livelihoods.
Policy-makers should consider establishing a national AI safety and governance council comprising experts from technology, industry, cybersecurity, law, economics, academia, and government.
Its purpose should not be to slow adoption, but to keep governance in step with expanding AI capabilities. It could identify high-risk applications, establish testing standards, recommend safeguards, and create a system for learning from serious AI-related failures.
Regulation should be risk-based rather than one-size-fits-all. A company using AI to summarize documents should not face the same requirements as a bank using it to assess loans. The greater the potential harm, the stronger the safeguards should be.
Businesses and industrialists should begin with applications where the benefits are clear and the risks manageable, such as document processing, internal research, demand forecasting, quality control, and routine customer service. They should be more cautious before giving AI greater autonomy over decisions involving customers' money, employees' livelihoods, or operational safety.
Humans should remain responsible for consequential decisions. Employees need training to recognize AI's limitations and verify its output. Cybersecurity should likewise be a leadership priority wherever AI systems connect to financial, customer or production systems.
Above all, business leaders should stop asking, “Where can we use AI?” and instead ask, “What problem are we trying to solve, and is AI the safest and most effective way to solve it?”
Adopting AI for its own sake can add cost and complexity without creating real value.
Keeping pace
Bangladesh cannot afford to miss the productivity gains AI offers. Our manufacturers, exporters, banks, and public institutions will increasingly need it to remain competitive. But we should not give increasingly autonomous systems responsibilities that our data, skills, infrastructure, and governance are not yet ready to support.
The goal is not risk-free AI -- that is unrealistic -- but an ecosystem in which risks are identified, assessed, and managed before they become crises. Bangladesh should embrace AI, but with its eyes open.
MM Shahidul Hassan is a distinguished professor at Eastern University, and former Vice Chancellor of East West University, Bangladesh.