Near-Term vs. Long-Term Risk

One of the most persistent sources of confusion in public discussions of AI risk is the failure to distinguish between harms that are already occurring and harms that might occur in the future. Both matter. But conflating them produces distorted priorities — either dismissing all AI risk because some future scenarios seem speculative, or treating near-term harms as less urgent than dramatic long-horizon concerns. Careful risk analysis requires separating what we know from what we project, and managing both with appropriate seriousness.

Near-Term Risks: Documented and Occurring Now

Near-term AI risks are harms that are either actively occurring or for which the causal mechanisms are well-established and the technical preconditions already exist. They do not require speculation about future AI capabilities. They are happening now, to real people, and they are the subject of active research, litigation, and policy debate. Bias and discrimination in algorithmic decision systems: AI systems used in hiring, credit scoring, pretrial risk assessment, and healthcare resource allocation have been documented to produce disparate outcomes by race, gender, age, and other protected characteristics. In 2019, a widely used algorithm for allocating additional healthcare to patients was found to systematically under-refer Black patients for supplemental care at the same level of medical need. The mechanism was not malice but a poor proxy: healthcare cost was used as a proxy for healthcare need, and because discrimination in healthcare access meant Black patients had historically been under-served, the algorithm reinforced that disparity. This is documented, quantified, and a genuine ongoing harm. Surveillance and privacy erosion: Facial recognition systems have been used by law enforcement with documented false-match rates that are significantly higher for darker-skinned women than for lighter-skinned men. Wrongful arrests based on facial recognition errors have been documented. Pervasive commercial data collection, combined with AI analysis, enables tracking and profiling of individuals at a detail and scale that is historically unprecedented. AI-generated disinformation and fraud: Deepfake audio has been used in documented CEO-fraud cases. Synthetic content has been used to generate non-consensual intimate imagery. AI-generated disinformation has appeared in documented election-related contexts. These are not hypothetical — they have case files.

Longer-Term Risks: Plausible but Uncertain

Longer-term AI risks involve scenarios whose mechanisms are plausible and whose technical preconditions are being actively developed, but whose probability, magnitude, and timeline are genuinely uncertain. They require extrapolation from current trends into a future that does not yet exist. Transformative automation and labor displacement: Current AI systems already automate substantial cognitive work. Whether AI will automate most cognitive tasks within a decade, several decades, or not at all is disputed. The labor market implications of different trajectories are radically different. This is a risk worth analyzing seriously, but honest analysis must acknowledge the genuine uncertainty about pace and extent. AI-enabled concentration of power: The economic dynamics of AI development create strong centralizing pressures. Whether this results in a small number of entities — companies or states — acquiring dominant and irreversible economic and political power is a plausible extrapolation, but it depends on regulatory responses, competitive dynamics, and technical developments that are highly uncertain. Advanced AI alignment failures: As AI systems become more capable and are deployed in more autonomous roles, ensuring that their behavior remains aligned with human values becomes more difficult and more consequential. The concern that sufficiently advanced AI systems might pursue objectives in ways that are harmful to humans — not out of malice but out of misalignment between their optimization targets and human values — is taken seriously by a significant fraction of researchers. The probability and timeline of such scenarios are among the most contested questions in the field. Serious researchers hold views ranging from 'this is the most important problem of our era' to 'this is not a near-term concern and focus on it distracts from immediate harms.'

Why Both Deserve Serious Attention

A common error in risk prioritization is treating near-term and long-term risks as competing: arguing that focus on existential long-horizon concerns distracts from near-term harms, or that focus on near-term harms is too short-sighted given longer-horizon concerns. This is a false tradeoff. Serious AI safety work addresses both, for complementary reasons. Near-term risks deserve attention because they are happening now. Every week of inaction on documented algorithmic discrimination or surveillance overreach has real costs to real people. Near-term risks are also more tractable: the mechanisms are understood, the victims can be identified, and the interventions are more concrete. Longer-term risks deserve attention because high-stakes developments require long lead times. If transformative AI capabilities arrive within a decade, the governance frameworks, technical safety research, and social institutions needed to manage them take years to develop. Waiting until the risk is imminent to begin preparation is waiting too long. The appropriate comparison is pandemic preparedness: we do not wait for a pandemic to begin before investing in vaccine research infrastructure, disease surveillance, and public health capacity. The two horizons also interact. The precedents set by how we handle near-term AI risks — whether we build robust oversight institutions, whether we establish norms of accountability for AI developers, whether we build international governance capacity — will shape whether longer-term risks are addressed well or poorly. Getting near-term governance right is foundational to long-term governance.