Why some people are stepping back from AI as concerns come from inside industry - explainer

SEPTEMBER 15, 2026 10:29The benefits of AI are becoming more apparent, but a growing backlash is raising an urgent question: Are companies deploying systems that can mimic people, analyze private data, write code, and act online before their developers have demonstrated that those systems can be safely controlled?For more stories from The Media Line go to themedialine.orgThe concern has become difficult to dismiss because it is coming from within the industry. Dario Amodei, chief executive of Anthropic; Sam Altman, chief executive of OpenAI; and Demis Hassabis, who leads Google DeepMind, have all recently backed a slower approach to developing the most advanced models. Their warnings surround cyberattacks, biological misuse, autonomous weapons, mass surveillance, and systems that may cause harm faster than people can understand or stop it. Public anxiety has risen as generative AI has become more visible. In a June 2026 survey, Pew Research Center found that 52% of Americans were more concerned than excited about the growing use of AI in daily life. Only 9% were more excited than concerned.People increasingly encounter AI in software, search engines, schoolwork, and public services without knowing what data are collected, whether the system is accurate, or who is accountable when it fails.Frontier models: the potential for AI hackersAmodei has made a clear call for restraint. In an essay published on Sept.12 entitled “We Must Pace the Frontier,” he wrote: “We must slow the pace at which we improve the capabilities of AI models.” CEO of Anthropic Dario Amodei, addresses the gathering at the AI Impact Summit, in New Delhi, India, February 19, 2026. (credit: REUTERS/Bhawika Chhabra)He is focused on frontier models - the most powerful systems, which can increasingly handle multistep tasks, write and run code, search for information, and use online tools with limited supervision. The concern is that their capabilities may improve faster than researchers’ ability to ensure that they consistently do what people intend.An AI system can follow an instruction literally while pursuing it through an unexpected or harmful route. This becomes more dangerous when AI is not merely answering questions but also carrying out tasks. An agent designed to solve a technical problem could be used to probe networks for weaknesses, write malicious code, send realistic scams, or adapt an attack when it is blocked. Human hackers are constrained by time and expertise. A capable AI system could operate continuously, test multiple strategies simultaneously, and create tailored material at scale. The danger is that people give it a goal, cannot foresee how it will pursue that goal, and recognize the harm it causes too late.Altman publicly backed Amodei’s central argument. “I agree with Dario that we need to pace the frontier,” he wrote. Companies, in his view, should not release a more powerful model merely because they have built one. They should first establish whether it can be misused, evade safeguards, assist dangerous activity, or cause failures that cannot be detected and contained. Altman also endorsed external scrutiny: “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”AI companies largely test their own systems and decide how much safety information to disclose. They also compete fiercely for users, investment, and influence. A firm that delays a launch may fear that a rival will go first.Can regulators verify AI safety tests?In a post on X/Twitter, Hassabis said Amodei’s proposals pointed “towards the right path forward,” but cautioned that “the details need working through.”Who decides that a model is too powerful to release? What does a meaningful safety test look like? How can regulators verify companies’ claims? International competition makes these questions harder - governments regard advanced AI as a source of economic, scientific, and military power, and safety can easily yield to the pressure of the race.Cybersecurity is among the clearest near-term risks. AI can help criminals and hostile states find software vulnerabilities, write malware, and generate convincing phishing messages. It can tailor scams to an individual using information taken from social media or data leaks and reproduce the style of a bank, employer, or government office. Voice-cloning technology adds another danger - a call that seems to come from a relative, manager, or official may not be genuine.AI magnifies the speed and scale of attacks. Hospitals, electricity providers, transport systems, and banks could face assaults generated faster than human security teams can investigate them.A second concern is biological and chemical misuse. AI can help legitimate researchers search scientific literature, understand molecules, and design experiments. But the same capabilities could help someone organize dangerous information, identify experimental routes, or seek ways around safety restrictions.AI cannot manufacture a biological weapon. Laboratories, equipment, materials, scientific skills, and testing remain essential. But sophisticated systems could lower the information barrier for a determined person or group that already has access to real-world resources. The danger lies in sustained technical assistance to someone pursuing a dangerous objective.AI as a tool to identify, target threats in warAI’s use in war poses an especially serious problem. Systems built to process information rapidly may move closer to decisions about whether people live or die.AI can analyze drone footage, satellite images, communications, and databases much more quickly than human analysts. This has the potential to help identify threats and prioritize intelligence. But critics warn that AI-assisted targeting can make lethal choices appear objective even when the data are incomplete, biased, or wrong.The issue is particularly relevant in the Middle East. Supporters say the technology helps manage vast quantities of intelligence. Critics and human-rights researchers argue that when human reviewers are expected to approve machine-generated recommendations rapidly, it could put civilians in unnecessary danger.Amodei has drawn a line against weapons that select and engage targets without meaningful human control. “Fully autonomous weapons … may prove critical for our national defense,” he wrote in Feb. 2026. “But today, frontier AI systems are simply not reliable enough to power fully autonomous weapons.”AI surveillance can reinforce discrimination instead of neutral judgementHe has also warned that “using these systems for mass domestic surveillance is incompatible with democratic values.” AI can combine information from cameras, phones, online behavior, purchases, and communications. Without strong legal limits, it could allow authorities to track political opponents, journalists, minority groups, and ordinary people at an unprecedented scale.AI can also reinforce discrimination while giving it the appearance of neutral, technical judgment. Systems trained on historical data can reproduce earlier bias in hiring, lending, welfare, policing, and healthcare. An algorithm may label someone high risk or unsuitable without disclosing the data or assumptions behind that judgment. This is why the European Union has prohibited some AI practices, including certain forms of biometric categorization.The danger is not only that AI makes mistakes, but that people may defer to erroneous conclusions because an AI recommendation can appear more objective than human judgment. This “automation bias” is especially dangerous in policing, medicine, welfare, immigration, and military decision-making, where a flawed recommendation can affect a person’s freedom, livelihood, or life.For many people, the most immediate danger is the erosion of trust in information. AI can generate realistic articles, photographs, audio, and video. It can imitate a person’s voice, fabricate a speech, or alter footage to suggest an event occurred.Spread of false material, disinformationDuring elections, war, or public emergencies, false material can spread before journalists and authorities can verify it. It can inflame ethnic or religious tensions, manipulate voters, discredit reporting, or cause panic. It also gives genuine wrongdoers an excuse to dismiss authentic evidence as artificial.In a Pew Research Center survey conducted in June 2025, 76% of Americans said it was extremely or very important to know whether pictures, video, and text were made by AI or by people. Yet 53% said they were not too or not at all confident that they could recognize AI-generated content. Once people believe that any recording may be fake, real evidence becomes easier to deny.A related threat is manipulation. AI can use personal data to predict what captures a person’s attention or to exploit a moment of vulnerability, then tailor messages accordingly. That may mean political propaganda aimed at a particular voter, gambling promotions directed at someone with an addiction, or loans and shopping offers targeted at a person in debt. The concern is not only false information but also systems designed to influence behavior in ways people may not recognize. The European Union’s Artificial Intelligence Act prohibits certain AI systems that manipulate people or exploit vulnerabilities linked to age, disability, or economic circumstances.Some people are stepping back from AI because it is moving into sensitive areas of life before clear limits are in place. The warnings from Amodei, Altman, and Hassabis demonstrate that the risks are serious enough that safety should not be treated as an afterthought.The issue is ultimately one of control. The real test will come when safety demands that a company delay a lucrative AI product, or when a government decides that a specific AI capability offers a military or intelligence advantage. The risk is that the most consequential choices about AI may be made in corporate boardrooms and closed government meetings.Follow us on Google