Reading view: this page was generated at build time from one reviewed JSON file. Enable JavaScript for search, filters, sorting, and pagination.
AI CAPABILITY TRACKER / RESEARCH SNAPSHOT
Explore what AI can do, what it still can't reliably do, and the evidence behind every breakthrough.
Latest review: Nov 08, 2026 · 66 capabilities in one reviewed JSON file
Counts are derived from the whole dataset and do not change when the list is filtered.
THE MOVING FRONTIER
Each bar counts the tasks with public evidence by that year. Read left to right: the line only goes up.
17 tasks with public evidence49 still waiting for a first public date
1 cumulative tasks with public evidence
2 cumulative tasks with public evidence
14 cumulative tasks with public evidence
Every Not Yet is a date that has not arrived yet. This is editorial framing, not a forecast: no timeline is promised, and no date is filled in ahead of evidence.
01 / OPEN QUESTIONS
Capabilities marked Not Yet or Partial. Newest review first, then title order, so the list stays stable between deployments.
Research assistance exists; human experts remain responsible for high-stakes decisions.
Requires permissions and audit trail.
Benchmarks and autonomy research do not establish indefinite universal reliability.
Classification can be automated; edge cases need monitoring.
Flags are not a substitute for legal advice.
OSWorld evaluation indicates significant remaining errors.
Plans are proposals, not proof of campaign performance.
Supplier terms must be verified and kept current.
Accuracy varies with noise and terminology.
No verified general-purpose zero-risk migration capability.
Discovery support is not equivalent to complete validated clinical development.
Narrative must be based on reconciled figures.
Knowledge retrieval and answer drafting work; escalation and source accuracy require review.
Human review required for fairness and legal compliance.
Personalization is possible; compliance and deliverability remain separate.
Existing deployments are geographically and operationally constrained.
Interpretations can omit runtime behavior.
Hypotheses are not verified causes.
OCR and extraction need validation before posting.
Verify quotations and paper relevance.
Accessibility and production implementation require testing.
Creative drafts need policy and brand review.
Answer keys need validation.
Do not invent product specifications.
Queries require permissions and execution checks.
Even high benchmark results do not imply universal mathematical correctness.
Universal perfect factual reliability has not been established.
Can assist with evidence review; not a universal substitute for peer review or experiments.
No model achieves perfect safety and accuracy across unrestricted cases.
Professional accountability cannot be inferred from model outputs.
No convincing evidence of reliable universal end-to-end business autonomy.
Stanford 2026 describes low real-household-task success relative to simulations.
General robust physical autonomy remains unproven.
Commercial terms require owner approval.
No reviewed evidence establishes flawless consistency end-to-end with zero human supervision.
Video generation exists; direction, editing and continuity still often require humans.
Perfect translation across unrestricted situations is not demonstrated.
Universal legal accuracy and accountable judgment are not established.
AI Index reports 50.6% top-model accuracy on ClockBench, versus 90.1% for humans.
Exceptions and financial controls require human review.
AI assistants do not inherently supply transaction integrity and compliance infrastructure.
Public signals can be summarized but may be incomplete.
Bounded agent workflows can be designed with review gates.
Bounded workflows are feasible; full process coverage and reliability vary.
Classification requires quality checks.
Sensitive personal data requires safeguards.
Does not replace incident response or controls.
Workflow design requires testing, permissions and rollback.
Coverage and assertions need independent review.
02 / THE FULL SNAPSHOT
Every reviewed record, including achieved ones. Achieved entries show the demonstrated tool, the publication date of the public evidence, and its source.
66 of 66 capabilities
Research assistance exists; human experts remain responsible for high-stakes decisions.
Requires permissions and audit trail.
Browse and synthesis available; accuracy depends on sources and review.
Historically demonstrated; check product availability separately
Benchmarks and autonomy research do not establish indefinite universal reliability.
Classification can be automated; edge cases need monitoring.
Flags are not a substitute for legal advice.
OSWorld evaluation indicates significant remaining errors.
Simulation and controlled tasks do not imply robust real-world generality.
Historically demonstrated; check product availability separately
Works on supported tasks; success is not guaranteed.
Historically demonstrated; check product availability separately
Plans are proposals, not proof of campaign performance.
Supplier terms must be verified and kept current.
Accuracy varies with noise and terminology.
Can produce editable outputs; financial and factual checks remain necessary.
Historically demonstrated; check product availability separately
No verified general-purpose zero-risk migration capability.
Discovery support is not equivalent to complete validated clinical development.
Narrative must be based on reconciled figures.
Knowledge retrieval and answer drafting work; escalation and source accuracy require review.
Human review required for fairness and legal compliance.
Personalization is possible; compliance and deliverability remain separate.
Existing deployments are geographically and operationally constrained.
Supports automatic multilingual dubbing; professional review may improve nuanced translations.
Historically demonstrated; check product availability separately
Can edit and refine imagery; precision and consistency are not guaranteed.
Historically demonstrated; check product availability separately
Interpretations can omit runtime behavior.
Hypotheses are not verified causes.
OCR and extraction need validation before posting.
Verify quotations and paper relevance.
Produces images from prompts; may still miss exact constraints.
Historically demonstrated; check product availability separately
Accessibility and production implementation require testing.
Creative drafts need policy and brand review.
Answer keys need validation.
Do not invent product specifications.
Documented May 2025; complex multi-scene continuity remains a limitation.
Historically demonstrated; check product availability separately
Queries require permissions and execution checks.
Even high benchmark results do not imply universal mathematical correctness.
Universal perfect factual reliability has not been established.
Can assist with evidence review; not a universal substitute for peer review or experiments.
No model achieves perfect safety and accuracy across unrestricted cases.
Professional accountability cannot be inferred from model outputs.
No convincing evidence of reliable universal end-to-end business autonomy.
Can use a virtual computer; sensitive actions may need confirmation.
Historically demonstrated; check product availability separately
Stanford 2026 describes low real-household-task success relative to simulations.
General robust physical autonomy remains unproven.
Predictions are computational, not proof of experimental validity.
Historically demonstrated; check product availability separately
Commercial terms require owner approval.
No reviewed evidence establishes flawless consistency end-to-end with zero human supervision.
Can research and synthesize sources, but references and claims require verification.
Historically demonstrated; check product availability separately
Video generation exists; direction, editing and continuity still often require humans.
Geofenced autonomous service, not all-road/all-weather autonomy.
Historically demonstrated; check product availability separately
Perfect translation across unrestricted situations is not demonstrated.
Universal legal accuracy and accountable judgment are not established.
AI Index reports 50.6% top-model accuracy on ClockBench, versus 90.1% for humans.
Exceptions and financial controls require human review.
Success is task- and evaluation-dependent.
Historically demonstrated; check product availability separately
AI assistants do not inherently supply transaction integrity and compliance infrastructure.
Public signals can be summarized but may be incomplete.
Bounded agent workflows can be designed with review gates.
Bounded workflows are feasible; full process coverage and reliability vary.
Specific documented competition problems; not unrestricted flawless mathematics.
Historically demonstrated; check product availability separately
Classification requires quality checks.
Structured extraction can work; input quality and validation remain important.
Historically demonstrated; check product availability separately
Sensitive personal data requires safeguards.
Does not replace incident response or controls.
Near-real-time translation across 70+ languages; domain accuracy varies.
Historically demonstrated; check product availability separately
Workflow design requires testing, permissions and rollback.
Can implement features and fixes in sandboxes; human review recommended.
Historically demonstrated; check product availability separately
Coverage and assertions need independent review.
METHODOLOGY / READ THIS FIRST