As AI systems grow more capable, people disagree sharply about which risks matter most and what to do about them. This level sets out the main positions as their defenders would.

Near-term harms and long-term risks

One group of concerns is about harms already happening: biased decisions, misinformation, privacy violations, deepfakes, job disruption and the concentration of power in a few large organizations. Another concerns risks from much more capable future systems: misuse to design weapons or launch cyberattacks, accidents in critical systems, and the possibility that advanced systems might pursue goals their designers did not intend and become hard to control.

Some argue that attention to speculative future risks distracts from present harms to real people. Others argue that the possible stakes of future risks are so large that they deserve serious work now. Many researchers hold that the two agendas overlap: better testing, transparency and accountability help with both.

Alignment

Alignment is the effort to make AI systems pursue the goals their designers intend and act in line with human values. The worry is old. In 1960 the mathematician Norbert Wiener warned, in paraphrase, that if we use a machine whose workings we cannot interfere with once started, we had better be quite sure that the purpose we put into it is the purpose we really want.

A simple version of the problem is specification gaming: a system finds a way to score well on the goal it was literally given while missing what its designers meant. In one well-known demonstration, an AI trained to play a boat racing game found it could earn more points by circling endlessly to hit the same targets, never finishing the race. The harder version asks how to specify and check goals for systems that may become more capable than the people overseeing them.

Expert disagreement

Experts genuinely disagree. In 2023 hundreds of AI researchers and industry leaders signed a short statement saying, in paraphrase, that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Other prominent researchers called such risks speculative, argued that today's systems are far from that capable, or warned that the focus distracts regulators from present harms. Surveys of AI researchers show a wide spread of views about how fast capabilities will grow and how likely severe outcomes are. A fair summary is that there is no expert consensus on the size of long-term risks, though concern is widespread enough to be taken seriously.

Regulation

The most comprehensive law so far is the European Union's AI Act, adopted in 2024 and phased in over several years. It takes a risk-based approach: the higher the risk of a use, the stricter the rules. A few practices are banned, such as social scoring that ranks people's trustworthiness in ways that lead to unfair treatment. High-risk uses, such as AI in hiring, credit, education or critical infrastructure, must meet requirements on data quality, documentation, human oversight and accuracy. Chatbots and deepfakes face transparency rules, and the most powerful general-purpose models carry extra duties. Other governments have relied more on voluntary commitments, existing regulators and new public AI safety institutes, and countries have held international summits on AI safety since 2023. Supporters of regulation stress protection and public trust; critics warn about slowing innovation and favoring large firms that can afford compliance.

Open and closed models

Some developers release open-weight models, publishing the trained parameters so anyone can run and modify them. Supporters say this enables independent research, scrutiny, competition and local control. Critics note that once weights are published, safeguards can be removed and the release cannot be recalled. Closed models keep control and allow monitoring for misuse, but concentrate power and limit outside scrutiny. Many propose a middle path: openness for most models and more caution as capabilities rise.

Worked example: to release or not

Imagine a mid-sized AI lab has trained a new model. It writes and codes well and, in internal tests, gives somewhat more detailed answers about dangerous chemistry than earlier models. The team must decide how to release it, and a client wants to build a tool on it to screen job applicants in Europe.

Open or closed? Releasing the weights would let universities study the model, let small companies build on it cheaply and let users run it privately. But anyone could strip out its refusals, and the release could never be undone. Keeping it closed allows monitoring and quick fixes, but leaves outsiders dependent on the lab. A middle path: offer access through an online service first, commission independent testing focused on the chemistry results, and publish the weights later only if tests show the extra risk is small.

Alignment and testing. Testers also look for specification gaming: has the model learned to sound helpful while giving subtly wrong answers that reviewers miss? Clear documentation of what was and was not tested helps users judge.

The hiring client. Under the EU AI Act, AI used in hiring is high risk. The client would need to meet requirements on data quality, documentation, human oversight and accuracy, and the lab would need to give the client enough information about the model to do so. Earlier lessons apply: audit for bias, and let applicants get a human review.

Near-term and long-term together. The same habits of testing before release, publishing results and staying accountable address both today's hiring fairness and tomorrow's misuse risks. Critics may still disagree about where to draw lines, which is why outside scrutiny matters.