In July, artificial-intelligence agents at OpenAI did something they were expressly meant not to do. During internal cybersecurity evaluations, models found ways around controls designed to keep them off the internet. They communicated through unauthorised channels, exploited weaknesses in shared infrastructure and eventually reached systems belonging to Hugging Face. One internal research model drove most of the activity. OpenAI's August investigation called the episode a "warning shot".
These were highly capable research models operating in specialised cyber tests with reduced safeguards. They did not suddenly develop human motives. Yet they crossed boundaries their designers had set, improvised ways around restrictions and pursued tasks in ways their operators had not intended. And it has generated significant and increasingly urgent debate since around the question: what happens when AI runs out of human control?
As technology leaders and governments in the US, China and elsewhere grapple with the pace of the AI race and the growing autonomy of artificial systems, another question sits much closer to home. What happens when we gradually stop exercising control ourselves?
AI now summarises what we might, and arguably should, have read, drafts what we might have written, recommends what we might have considered and increasingly proposes decisions we might once have made ourselves. Much of this is useful, but the risk grows when this form of assistance becomes substitution and convenience becomes a habit that means surrendering judgement.
Research in 2026 is sharpening that concern. A study of young adults comparing AI-assisted and manual problem-solving found evidence of cognitive offloading: AI users could perform tasks with less mental effort, while some delegated substantial parts of the reasoning process. A July Scientific Reports study likewise found that patterns of AI use were associated with differences in self-rated critical thinking. One experiment deliberately gave participants advice that was correct only half the time. People with more positive attitudes towards AI relied more on the faulty AI guidance and became worse at distinguishing real from synthetic faces.
A September paper in AI & Society describes a broader shift: from decision support toward "decision substitution". As systems increasingly shape what humans see, validate and choose, authority migrates from human judgement to algorithmic output. This is agency decay or, in plain language - the gradual weakening of our ability to make our own decisions. We experiment. We integrate. We rely. At the far end, reliance can harden into addiction. The change is seemingly undramatic. We simply exercise the mental muscle less often.
Before assuming that smarter machines will somehow become wiser machines, our own record deserves a glance. Before assuming that smarter machines will somehow become wiser machines, our own record deserves a glance.
Since 1990, humanity has cleared 489 million hectares of forest—an area almost one and a half times the size of India. We now produce more than 400 million tonnes of plastic each year, while the equivalent of a garbage truck of plastic enters the ocean every minute. Human activity has pushed around one million animal and plant species towards extinction, with extinction rates tens to hundreds of times higher than the average over the past ten million years. In 2025, atmospheric CO₂ reached around 53% above pre-industrial levels, after decades of accumulating evidence about the causes and consequences of greenhouse-gas emissions.
These are not failures of intelligence. They show what can happen when intelligence expands faster than wisdom, restraint and collective foresight. We became extraordinarily good at extracting, producing and optimising, while remaining far less accomplished at recognising limits. That distinction matters for AI. Greater intelligence expands what an agent can do. It does not determine what it should do, whom it should serve, or when it should stop.
Human beings have accumulated extraordinary scientific knowledge while pushing seven of nine planetary boundaries beyond their assessed safe limits. Knowledge alone does not make individuals more kind nor more responsible in their behaviour.
AI arises from this human record. It learns from human-produced data: our discoveries and prejudices, generosity and cruelty, insight and misinformation. Humans choose their goals. Human institutions decide what gets rewarded. Companies decide whether an algorithm should maximise attention, efficiency, revenue or wellbeing. Governments determine where machines may act and where human responsibility must remain.
"Garbage in, garbage out" was one of computing's earliest lessons. In the age of AI, we need a more consequential version: values in, values out.
Values live in objectives, incentives, data, permissions and limits. Tell an agent to achieve a result without adequately defining acceptable means, and capability can become a liability. Ask a system to maximise engagement, and it may learn that outrage keeps people watching. Reward speed while making human review inconvenient, and review will gradually disappear.
The same mechanism applies to us. Every time we accept an answer we have not considered, forward a claim we have not checked or outsource a judgement we should own, we help contribute toward the establishing of a social norm about where authority belongs.
Malaysia's National AI Action Plan 2026-2030 sets out a vision of a "value-driven AI Nation". That phrase should be taken literally. The country is building AI capability, infrastructure and institutions while the rules remain malleable. This creates an opportunity to ask before adoption becomes automatic: does this technology strengthen human and planetary wellbeing, including the agency of the people using it?
That is the purpose of ProSocial AI. It asks us to judge technology by what it enables across four dimensions: Purpose, People, Prosperity and Planet. Doing so requires Double Literacy: sufficient technological understanding to use AI intelligently, coupled with the human capacity to question, reason critically, exercise compassion and make choices with an awareness of their wider consequences.
For individuals, this offers a practical discipline: think before prompting. Check before accepting. Ask what evidence, perspective or consequence may be missing. Keep consequential decisions in human hands, and understand the reasoning behind them. For schools and universities, the task is to teach students to think and work effectively with AI while preserving their ability to reason, create and decide independently of it. For organisations, it means assessing which human capacities a system strengthens, which it weakens and which it displaces. For government, it means creating incentives and ccountability mechanisms that reward outcomes serving humanity and the planet, while giving businesses clear, workable guardrails—and meaningful reasons to respect them.
The technology of tomorrow will inherit more from us than our data. It will inherit our priorities, incentives and habits.
We cannot expect it to become better than the humans of today unless we become more deliberate about what we put into it.
Garbage in, garbage out described the old computer age. Values in, values out may decide the next one.
Dr Cornelia C Walther is associate professor at Sunway University’s Institute for Global Strategy and Competitiveness and senior fellow at the Sunway Centre for Planetary Health. She advises UNFPA on Hybrid Intelligence and is a senior fellow at Harvard and the Wharton School.