AI Governance Gaps Enable Bias and Misinformation
Computer science conferences publish papers that claim new models can predict creditworthiness or detect fake news with near-perfect accuracy. But when these systems leave the lab and enter a lending platform or social media feed, the promise usually cracks. The gap between what researchers demonstrate and what ends up in real products is not just a technical delay — it is the opening through which deception enters. Malaysia’s proposed Artificial Intelligence Governance Act, as debated by experts like Dr Ainuddin, is trying to close that gap before it becomes a canyon.
The Misrepresentation of What AI Actually Does
Modern machine learning does not reason the way a human does. It finds statistical patterns in data that are often spurious — a correlation between loan defaults and the time of day a form was submitted, for instance, or between health outcomes and a zip code that proxies for income. When a bank deploys such a model, it appears to be making rational decisions. In reality, it is reproducing hidden biases that the system designers may not even know exist. This is the first layer of deception: the AI looks smart, but its intelligence is a mirror that reflects every flaw in the training data. Dr Ainuddin points out that the current risk framework in the Bill defines harm only in rigid terms like death or physical injury. Ainuddin2024 That misses the quiet erosion that happens when a hiring algorithm consistently filters out candidates from a certain background — no one dies, but opportunity is silently stolen. This deception is that the system appears objective while it quietly entrenches old prejudices.
Algorithmic Bias as a Slow Erosion of Trust
Hiring software that learns from historical CVs can penalise women for career breaks or penalise candidates from less prestigious universities because past hires all came from elite schools. The model does not know it is being unfair; it simply maximises a correlation that happened to work before. Dr Ainuddin warns that such bias does not directly injure someone the way a faulty machine would, but it erodes trust and fairness over time. This is a form of deception because the system does not announce its prejudice. A human recruiter might explicitly explain why a candidate was rejected, but an AI gives a score or a “no” with no reasoning. The user — the job applicant — is left with a feeling of injustice but no clear target. The Bill’s reliance on the Personal Data Protection Act is insufficient, as Dr Ainuddin notes. The PDPA guards the door, controlling who collects data, but it does not question what the AI does inside. A loan applicant might consent to data sharing, but if the AI denies the loan based on a biased pattern, the deception is complete: consent was given, but fairness was not.
The Loophole in National Security Exemptions
Governments around the world exempt national security systems from public oversight, and Malaysia’s Bill follows that pattern. Dr Ainuddin warns that “national security” can be a wide umbrella. A surveillance system using facial recognition might claim to be protecting citizens, but if its algorithm misidentifies faces more often for certain ethnicities — a phenomenon studied in computer vision research. The people affected never know they are being watched, let alone that the system is biased against them. This is a deep deception: the system operates in secrecy under the banner of protection, while its flaws remain invisible. Lawyer Thulasy Suppiah adds that exemptions should not remove accountability. Suppiah2024 An exemption from public disclosure should not be an exemption from responsible governance. The deception is that the public trusts the shield of national security, but the shield may have a hidden crack.
Correlation Is Not Causation, and AI Companies Know It
The boundary between correlation and causation is the structural fault line in most AI research. A model trained on hospital records might learn spurious correlations that could lead to harmful recommendations. The model confuses correlation with causation — ambulance patients are more urgent, not harmed by the ride. When such a system informs clinical decisions, it deceives doctors into making harmful choices. Dr Ainuddin’s concern about the Bill not addressing automated decision-making directly relates to this. The law must demand that systems explain their reasoning, not just produce a number. Thulasy Suppiah insists that people should know when AI was used for decisions about employment, loans, healthcare, or access to public services. That transparency is the antidote to the hidden cause-effect confusion. Without it, every AI output is potentially a lie disguised as a fact.
The Missing Guardrails for Continuous Learning
Modern AI is not a static machine. It learns after deployment through fine-tuning and third-party plug-ins. Dr Ainuddin points out that the Bill treats AI as if it stays the same after installation, but a model that starts with fair lending criteria can drift into bias as it absorbs new data. A hiring system might begin by ranking candidates by generic skills, but after a year of self-learning, it might unconsciously penalise non-native English speakers because they tend to score lower on a language test that has nothing to do with job performance. This deception is gradual: the system was once fair, so the users assume it still is. The Bill needs dedicated safeguards that require monitoring of algorithmic behaviour over time, not just a one-time approval. Without that, the AI will deceive by drift.
Personal Use Exemptions as a Backdoor
The
Bill exempts personal use of AI. Dr Ainuddin uses the example of sorting holiday photos — regulators do not need to step in there. But Thulasy Suppiah warns that “personal use” must be narrowly defined. A small business could claim they are using AI for personal family tasks to avoid regulation, while actually deploying it to screen rental applications. That would be a classic deception: the appearance of a hobby hides systematic decision-making that affects other people. The line between personal and commercial use is blurring as AI tools become cheap and powerful. A landlord using an off-the-shelf AI to decide which tenant to rent to is making a societal impact, even if they only have one property. The law must close this loophole, or the exemption becomes a licence to deceive.
The Last Open Variable: Political Will to Write Guardrails
Every expert in the parliamentary hearings agrees on the technical fixes: clear definitions of harm, mandatory transparency, a right to human review, and explicit bans on algorithmic bias for critical decisions. But the last variable that determines success or failure is not technical. It is political. Thulasy Suppiah stresses that people should have access to an appropriate explanation and a human review. That requires courts or agencies to enforce it. Dr Ainuddin calls for a clearer definition of harm that includes social erosion, not just physical injury. That requires legislators to listen. The gap between research and policy is not a technical gap; it is a gap in will. If the Bill passes with broad, vague exemptions and no binding oversight of algorithmic outcomes, then the AI systems it fails to regulate will quietly deceive the public for years. The question is whether lawmakers will close that last open variable or leave it gaping.