What AI can actually do, domain by domain.

AI already reads medical images alongside radiologists, flagging likely tumors, fractures, and diabetic retinopathy often as accurately as a specialist, and diffusion-based segmentation tools now outline organs and lesions in seconds rather than the hours a manual trace takes. Beyond imaging, large language models increasingly draft clinical notes, summarize a patient’s history before a visit, and help clinicians keep up with a research literature growing far faster than any one person can read.

What AI is not yet trusted to do, and rightly so, is make unsupervised final diagnostic or treatment decisions: regulators in essentially every major market require a human clinician in the loop, and AI systems remain prone to confidently wrong answers on cases unlike their training data.

Language models can now solve International Mathematical Olympiad problems at a medal-winning level, and “autoformalization” tools translate informal, human-written proofs into a fully machine-checkable formal language, letting a computer verify every logical step of a proof with total rigor. Mathematicians increasingly use AI as a tireless collaborator for exploring conjectures, searching for counterexamples, and formalizing textbook results into machine-verified libraries.

AI-generated proofs of genuinely novel, research-frontier theorems remain rare; today’s systems are far better at proving things a mathematician already strongly suspects are true than at proposing an entirely new true conjecture worth proving.

Generative models now propose candidate crystal structures and predict their stability computationally, screening millions of hypothetical materials before a single one ever touches a lab bench. Google DeepMind’s materials project reported identifying hundreds of thousands of potentially stable new materials this way, a search space that would take conventional trial-and-error many decades to cover.

The bottleneck has shifted from proposing candidates to synthesizing and testing them physically, since a computationally “stable” material still has to actually be manufacturable at a useful cost and scale, a step AI does not yet meaningfully accelerate.

DeepMind’s AlphaFold, which predicts a protein’s 3D shape directly from its sequence, won its creators the 2024 Nobel Prize in Chemistry and has been used to predict the structure of essentially every protein known to science, a problem that used to take years of laboratory crystallography per protein. Generative models now go further, designing entirely new candidate drug molecules and even new proteins with a specified function from scratch.

Several AI-discovered drug candidates have entered human clinical trials, and a handful have reached later trial stages, but as of today none has yet completed the full regulatory approval pipeline purely on the strength of an AI-originated design.

AI tools now draft contracts, summarize lengthy discovery documents, and search case law dramatically faster than a junior associate could, and several jurisdictions are piloting AI-assisted case-triage to help clear enormous court backlogs.

The same tools have also produced well-publicized, embarrassing failures, filings citing entirely fabricated court cases, precisely because a language model will confidently invent a plausible-sounding citation rather than admit it does not know one. The EU’s AI Act specifically flags AI systems used in the administration of justice as “high-risk,” subject to extra scrutiny.

Governments increasingly use AI to model the likely effects of a proposed policy, summarize enormous volumes of public comment on draft regulations, and translate government services into more languages than human staff could cover alone.

The same generative capability that drafts a helpful policy summary can just as easily produce a convincing deepfake of a candidate, which is exactly why the EU, the US, and the OECD’s AI governance observatory have all made election-related synthetic media a top near-term policy priority.

AI already assists forensic analysis, digital evidence triage, and pattern detection across large caseloads, helping investigators surface connections across cases far faster than manual cross-referencing.

Predictive policing and facial recognition remain the field’s most contested uses: independent government testing has repeatedly found meaningfully higher error rates for some demographic groups than others, driving outright bans or moratoria on facial recognition in policing in a growing number of cities and countries.

On the defensive side, AI already triages security alerts, detects anomalous network behavior, and helps patch vulnerabilities faster than human teams working alone. Offensively, the same underlying language-model capability can write malicious code and craft highly convincing phishing messages at a speed and scale no human attacker could match unaided.

National frameworks like NIST’s AI Risk Management Framework in the United States and the EU cybersecurity agency ENISA’s guidance now explicitly treat AI-specific risk as a distinct category of national security planning, not merely an extension of ordinary cybersecurity.