Perplexity Not Citing Sources? 8 Fixes That Work

Perplexity not citing sources is not a cosmetic inconvenience: in published research on generative search engines, only 51.5% of generated sentences were fully supported by citations, which means the missing link is often a warning that the evidence chain has already become fragile. I have seen the same failure pattern in live editorial work: an answer looks polished, numbered footnotes appear in one run, then the next run either hides them, maps them to the wrong page or synthesises a confident paragraph from several sources without attaching a source to the exact claim.

The practical answer is to stop treating Perplexity citations as decoration and start treating them as a claim-control system. This article explains why Perplexity sometimes produces answers without visible or correct citations, which settings and prompts change that behaviour, how Pro Search, Spaces, files, connectors and the Sonar API alter the workflow, and what a repeatable verification process looks like in 2026. It also separates confirmed pricing and plan caps from unverified assumptions, because citation reliability is partly a product design problem and partly an operational problem. By the end, readers should be able to force stronger citations, identify weak or mis-mapped links, build an API parser that does not drop sources, and decide when Perplexity is useful as a research assistant rather than an authority.

Why Perplexity Not Citing Sources Happens

The phrase “Perplexity not citing sources” covers several different failures. The simplest is prompt ambiguity. If a user asks for a broad explanation, the system may optimise for a fluent answer and attach only a small source set, especially when the answer contains background knowledge that does not map cleanly to one page. A tighter prompt that asks for “a citation after every factual claim” changes the task from explanation to auditable evidence extraction.

The second cause is mode mismatch. Perplexity is not a single surface. The consumer web interface, Pro Search, Research, Spaces, file uploads, connectors, Sonar API, Search API and third-party clients do not expose citations in identical ways. Perplexity’s API documentation distinguishes raw Search API results from Sonar answers with built-in citations, while the Agent API citation guide says numbered references can arrive in text while corresponding source URLs appear in a structured search results item.

The third cause is source parsing. Some pages are dynamic, paywalled, blocked, duplicated, stale or weakly structured. A model can retrieve enough text to answer part of the question but fail to attach a clean source to every sentence. That is why a citation can be absent even when the answer is broadly true. The output is a synthesis, not a database record.

The fourth cause is client rendering. Community issue reports from 2025 described cases where footnote markers appeared but citation links were not displayed because the third-party client did not parse the citation object. That is not the same as Perplexity failing to retrieve sources. It is a display-layer failure.

Failure Mode	What the User Sees	Most Likely Cause	Best First Fix
No citations at all	A fluent answer with no source badges	Prompt too broad, wrong mode or general synthesis	Ask for inline citations after every factual sentence
Footnotes without links	Numbers such as [1] or [2] but no clickable source	Client or API parser dropped the citation object	Inspect raw response fields and parse citations separately
Wrong citation opens	A link appears but does not support the sentence	Citation mapping drift or weak retrieval	Open the page and mark the claim as unsupported or partial
Few sources for a complex answer	A long paragraph supported by one or two links	Over-compression across multiple pages	Run a source-first pass before drafting
Sources missing after regeneration	One answer cites, the next does not	Non-deterministic retrieval and source ranking	Save source snapshots and ask for a claim ledger

That distinction matters because each failure needs a different repair. A user cannot fix an API parsing bug with a better research prompt, and a developer cannot fix weak source quality by merely styling links. The operating rule is simple: identify whether the problem is retrieval, generation, mapping or rendering before changing the workflow.

Prompt Phrasing and Mode Choices That Change Citations

Perplexity is most reliable when the prompt states the evidence contract before the answer begins. During our 2026 evaluation, the strongest prompt pattern was not clever wording. It was restrictive wording: “Use only verifiable sources, place a citation after every factual sentence, and say evidence is insufficient where a claim cannot be sourced.” That forces the model to make unsupported gaps visible.

Mode also matters. A quick web answer is useful for orientation, but high-risk work needs a source-first sequence. In one pass, ask Perplexity to return only sources with title, publisher, author, date and relevance. In the second pass, ask it to extract claims from those sources. In the third pass, ask it to write using only the claim ledger. That approach fits the broader Perplexity AI prompting guide approach because it puts constraints before style.

Focus choices affect source quality as much as citation quantity. If the question is academic, ask for peer-reviewed papers, DOI-bearing articles, university repositories or official research reports. If the question is commercial, ask for official pricing pages, help centre articles and product documentation. If the question is legal, medical or financial, use Perplexity only for source discovery and keep final interpretation with a qualified professional.

The strongest practical prompt is: “Do not answer yet. First list the sources you would use, with publication date and why each source is relevant. After I approve the sources, write the answer with a citation after every factual claim. If a claim is unsupported, label it unsupported rather than guessing.” This reduces orphan claims because the model must reveal the evidence set before composing the prose.

Users who already rely on the best features of Perplexity AI should treat citations as the starting point of verification, not the end. A better source badge helps only if the reader opens it, checks the date and compares the exact wording against the claim.

Perplexity Not Citing Sources in the API and UI

The API adds a second layer to the citation problem: structured data can exist in the response while the interface fails to show it. Perplexity’s streaming citation guide describes a design where the text can include numbered references and the corresponding source URLs arrive in a separate structured output item. If a third-party client renders only the text, the user can see footnote numbers without links.

That exact pattern has appeared in public developer reports. A June 2025 Chatbox issue described Perplexity responses with footnote references but missing citation links, and the report noted that the API provided citations that needed to be parsed. Similar reports from other clients and gateways described citation objects being dropped by OpenAI-compatible layers. The article context is important: the bug may sit in the wrapper, not in the model.

Incorrect mapping is a different problem. A citation number can point to the wrong item in the search results array, especially in complex, domain-filtered or Pro Search-like workflows. In automated research systems, this is more dangerous than a missing citation because it creates false confidence. The link is clickable, but it sends the reviewer to a source that does not support the sentence.

For API teams, the answer is to store four objects separately: generated text, citation list, search results, and a claim ledger that maps each factual sentence to a source ID. Do not rely on rendered markdown alone. A safe client should validate that every citation marker in the text maps to a resolvable URL and that every critical claim has at least one source record.

This is where a broader Perplexity hallucination fix becomes a systems problem. The safest architecture makes source retrieval, answer generation, citation rendering and human verification separate stages, each with logs.

Layer	Required Data to Store	What Can Go Wrong	Audit Control
Retrieval	Title, URL, snippet, date, last updated and retrieval time	Relevant pages are missed or weak pages rank highly	Log all source candidates before drafting
Generation	Answer text and citation markers	The model writes a claim that no source supports	Require a citation after every factual sentence
Mapping	Citation marker to source record relationship	Marker [3] opens the wrong page	Validate indexes and resolve URLs before display
Rendering	Hyperlinks, source cards and visible footnotes	The UI drops citations even though the API returned them	Test the client against raw API payloads
Review	Claim status and reviewer notes	A weak citation passes because it looks official	Mark claims confirmed, partial, unsupported or disputed

Pricing, Plans and Limits That Affect Citation Work

Citation reliability is partly shaped by plan limits. A free user can test prompts, but documented Perplexity plan information lists Free at 3 Pro Searches per day and 1 Research query per month. That makes repeated citation regeneration harder. Paid and enterprise tiers provide more capacity, but they also introduce governance, file and API budgeting questions.

The current public enterprise pricing page shows lower annualised monthly figures when billed yearly, while the Help Center describes Enterprise Pro as starting at $40 per month or $400 per year per seat. It also lists Enterprise Pro at 400 Pro Searches per week and 50 Research queries per month, while Enterprise Max rises to 4,000 Pro Searches per week and 500 Research queries per month. These figures should be checked at procurement time because regional pricing, app-store billing and contract terms can vary.

This matters for citation workflows because verification consumes queries. A source-first pass, claim extraction pass, drafting pass and citation audit pass can use several advanced searches for one deliverable. A newsroom, law firm or analyst team may hit limits faster than a casual user because it is not asking more questions for curiosity. It is performing evidence control.

Enterprise plans also change privacy and administration. Perplexity Help Center material says Enterprise Pro and Enterprise Max data is never used for training and that admins can enforce strict privacy and organisation controls. That is relevant when uploaded files, connectors and internal repositories become part of the citation context.

For buyers comparing seat economics, the Perplexity enterprise pricing decision should not be treated as a generic upgrade. The practical question is which users need higher research capacity, governance features or larger file contexts to verify sources properly.

Plan or Surface	Current Published Price Signal	Citation Capacity Signal	Limit or Caveat
Free	$0	3 Pro Searches per day and 1 Research query per month	Useful for light checks, weak for repeated verification
Pro	$17 per month when billed annually on enterprise pricing page, commonly presented as $20 monthly for consumer billing	Weekly limits for advanced use	Region and billing route can change the displayed price
Enterprise Pro	Starts at $40 per month or $400 per year per seat	400 Pro Searches weekly and 50 Research queries monthly	API usage is billed separately
Enterprise Max	Enterprise pricing page shows $271 monthly per seat when billed annually, Help Center lists highest enterprise limits	4,000 Pro Searches weekly and 500 Research queries monthly	Best for high-volume research and larger file workflows
Sonar API	Pay as you go	Citations available in web-grounded answers	Token, request, search, citation and reasoning costs vary by model
Search API	$5 per 1,000 requests	Returns raw ranked web results before generation	No prose answer unless paired with generation

Features, Integrations and Technical Specs That Matter

A complete citation workflow depends on more than asking for links. Perplexity’s documented product surfaces include the web answer engine, Pro Search, Research and Deep Research, Spaces, file uploads, connectors, Perplexity Computer, Sonar API, Search API, Agent API and Embeddings API. Each surface changes what can be cited, how sources are selected and how a reviewer should audit the result.

The Search API is the cleanest retrieval layer because it returns a structured result set with fields such as title, URL, snippet, date and last updated. It is useful when a developer wants to control which sources enter a downstream summarisation pipeline. Sonar is the web-grounded answer layer. It produces prose with citations and supports streaming, conversation context and model-specific capabilities.

Media and attachment handling introduces hard limits. Official documentation for Perplexity API media processing lists supported formats including PDF, DOC, DOCX, TXT and RTF, with a 50 MB per-file size limit and a maximum of 30 files per request for API media workflows. Spaces documentation lists persistent file limits by plan. When documents are scanned, image-heavy, table-heavy or poorly structured, citation accuracy can still suffer even if the file is accepted.

Connectors add another source of complexity. Perplexity’s 2026 changelog said Pro, Max and Enterprise subscribers could connect external tools and data sources through Model Context Protocol, with OAuth, API key or open authentication options. That can improve relevance, but it also raises stale-data and permission questions. If a connector pulls outdated project notes, the model can cite internal context that no longer reflects the current truth.

The safest comparison is captured in our Perplexity vs Google for research analysis: Perplexity is fast at synthesis and source discovery, while traditional search remains valuable for manual inspection, source diversity and checking whether the cited page is the original authority.

Feature or API	Technical Role	Citation Relevance	Known Constraint
Pro Search	Multi-step research in the app	Finds deeper source sets for complex questions	Plan limits affect repeated verification
Research and Deep Research	Longer autonomous investigation	Useful for broad source discovery	Long reports need sentence-level citation audits
Spaces	Persistent project context with files and instructions	Narrows the source pool for recurring work	Old files and connectors can pollute context
File Uploads	Ground answers in documents	Can cite or summarise user-provided files	50 MB and file-count limits vary by surface
Search API	Raw ranked web retrieval	Best for source-first pipelines	Requires a separate generation layer
Sonar API	Web-grounded answer generation	Returns cited prose	Citation parsing must be implemented correctly
Agent API	Multi-provider orchestration and tools	Can return structured search results	Tool and model behaviour varies by preset
Connectors and MCP	Bring external work tools into context	Can ground answers in workplace data	Permissions and stale source control become critical

Eight Fixes That Work for Missing or Weak Citations

The fixes below are deliberately operational. They do not assume that Perplexity will become perfect. They make citation failure easier to see, easier to reproduce and easier to correct.

1. Ask for citations before the answer begins. Use wording such as “place a citation after every factual sentence” rather than “include sources”.
2. Run a source-first pass. Ask for sources only, with publisher, date and relevance, before any prose is written.
3. Specify acceptable source classes. Pricing needs official vendor pages, statistics need primary reports, quotes need original interviews or reputable news reports.
4. Use the right mode. Choose Research, Pro Search, Academic-style prompts or API retrieval depending on whether you need speed, scholarly sourcing or structured data.
5. Regenerate with a narrower scope. If a broad answer lacks citations, split the query into smaller factual questions.
6. Ask “which source supports this sentence?” for unsupported claims. This forces a claim-level check instead of a general source list.
7. Open every key citation manually. A link proves retrieval, not support for the exact sentence.
8. In API clients, parse citation and search result objects separately. Do not rely on markdown footnotes as the only source record.

These steps work because they change the task shape. Instead of asking Perplexity to be “accurate”, they require it to show evidence, preserve uncertainty and separate source discovery from drafting. The most important habit is opening the source page. In our hands-on testing, most dangerous failures were not bizarre hallucinations. They were subtle source overreach, where the page supported one part of a sentence but not the conclusion attached to it.

When Perplexity gives a wrong or unsupported answer, use the same discipline described in Perplexity AI gives wrong answers: isolate the factual claim, identify the required source class, check the cited page, then either remove, revise or caveat the claim.

A Citation Verification Workflow for Researchers and Editors

A practical verification workflow needs to fit a real deadline. The method below is designed for journalists, analysts, marketers, academic researchers and B2B teams that use Perplexity to accelerate research but still need defensible sourcing.

First, split the answer into factual claims. A single sentence can contain multiple claims. “Perplexity Enterprise Pro costs $40 per seat and includes 400 weekly Pro Searches” contains a pricing claim, a plan claim and a usage-limit claim. Each claim needs a source. Second, assign a source class. Pricing claims require an official Perplexity page. News claims require a reputable report with a date and named publication. Research claims require a paper, dataset, report or official study.

Third, open the cited source and check exact support. Do not trust snippets. Confirm that the source says the same thing, that it is current enough and that it is not quoting another weak source. Fourth, mark the claim status: confirmed, partially supported, unsupported, contradicted or specialist review required. Fifth, rewrite the answer so uncertainty appears in the prose.

The point is not to slow down every casual search. The point is to reserve heavy verification for claims that carry consequences: prices, dates, limits, quotes, medical statements, legal statements, financial figures, product specifications and named allegations. Those are the nouns and numbers that turn an AI answer from useful to risky.

Claim Status	Definition	Editorial Action
Confirmed	The cited source directly supports the exact claim	Keep the claim and citation
Partially Supported	The source supports part of the claim but not the full wording	Rewrite narrowly or add a caveat
Unsupported	The page is relevant but does not prove the claim	Remove the claim or find a stronger source
Contradicted	The source or another authoritative source says the opposite	Correct the claim and document the conflict
Specialist Review	The claim is technical, legal, medical or financial	Route to a qualified reviewer before publication

This is also how teams should handle uploaded PDFs and large files. Ask Perplexity to cite a page, section or quoted phrase where possible, then inspect the document yourself. For citation-heavy academic work, the related how to cite Perplexity AI guidance is useful, but remember that citing the tool is not a substitute for citing the primary sources it surfaced.

Benchmarks and News Studies That Should Shape Trust

The evidence base argues for calibrated trust, not rejection. Liu, Zhang and Liang’s generative search study found that generated answers were fluent and informative but often under-supported, with only 51.5% of generated sentences fully supported by citations and 74.5% of citations supporting their associated sentence. That finding explains why a visible citation badge should be treated as a lead to inspect rather than a guarantee.

News research points in the same direction. The EBU and BBC study of more than 3,000 AI assistant responses across 14 languages found that 45% of responses had at least one significant issue, and sourcing was the largest significant issue category. Jean Philip De Tender, EBU Media Director and Deputy Director General, warned that when people do not know what to trust, they may end up trusting nothing at all.

The earlier BBC study on current affairs found serious distortions across AI assistants. Deborah Turness, then BBC News chief executive, warned that “Gen AI tools are playing with fire” and could undermine fragile faith in facts. That quote belongs in any discussion of AI citations because source links can create a false impression that the answer has already been professionally checked.

Perplexity’s growth raises the stakes. TechCrunch reported that CEO Aravind Srinivas said Perplexity handled 780 million queries in May 2025 and was growing more than 20% month on month. He also said the company could reach a billion queries a week if that growth held. At that scale, even small citation misunderstandings become common user experiences.

Stanford HAI’s 2026 AI Index adds a broader risk signal: documented AI incidents rose from 233 in 2024 to 362 in 2025, while hallucination rates across 26 top models ranged widely depending on test design. The exact number will vary by benchmark, but the decision rule is stable. Treat citations as evidence to verify, not proof to inherit.

Implementation Workflow for API Teams

Developers building Perplexity into products should design citation handling before launch, not after users complain. The minimum architecture has five stages: retrieve sources, generate answer, extract claims, validate citations and store an audit record. Each stage should be observable. If a citation disappears, the team needs to know whether retrieval failed, the model omitted a marker, the parser dropped a field or the UI hid the link.

A source-first pipeline can use Search API to gather candidate pages, then pass approved sources into Sonar or another generation layer. The application should store title, URL, snippet, date, last updated, retrieval timestamp and source rank. It should also store the model, prompt, parameters and response ID. Without these records, later audits cannot reconstruct why a particular source was trusted.

Next, use a claim ledger. Split the answer into claims, assign source IDs and mark evidence status. For high-stakes content, add a human review queue for unsupported or partially supported claims. This is not over-engineering. It is the difference between a search demo and a compliance-ready research workflow.

Rate limits and costs should shape design. Perplexity API pricing lists separate costs for Search API requests, Sonar tokens, request fees by context size, citation tokens for Deep Research, search query charges and reasoning tokens. A naive verification loop that re-runs the full query several times can become expensive. Cache source candidates, reuse verified sources where appropriate and avoid using Deep Research where a raw Search API call is enough.

For teams experimenting with Labs, Pro Search and multi-model review, the Perplexity Labs vs Pro Search distinction matters. Use the faster answer layer for ordinary source discovery and reserve heavier modes for complex questions where the extra cost and time can be justified by better evidence coverage.

Performance Bottlenecks and User Constraints

The main bottleneck is not always model intelligence. In citation work, the limiting factor is often source quality. A weak page can be retrieved, summarised and cited perfectly while still being the wrong source for the claim. Another common bottleneck is page structure. Dynamic pages, inaccessible PDFs, tables embedded as images, regional pricing pages and blocked news sites can prevent clean extraction.

Large context can also reduce trust if it is unmanaged. Spaces and connectors help Perplexity see relevant files, but they can mix old drafts, duplicated exports and conflicting versions. The more context a Space contains, the more important metadata becomes. File date, owner, source priority and version status should be stated in the prompt. Otherwise the model may cite an old internal memo as if it were the current policy.

User behaviour is the quietest constraint. Many people ask for citations, receive them and never click them. That collapses the entire trust model. Citation workflows work only when users accept that source verification is part of the task. The fastest way to improve output quality is to verify nouns and numbers first: prices, dates, names, quotes, product limits, statistics, laws and URLs.

Commercial limits can also nudge behaviour. If a user has few advanced searches left, they may stop before running the source-first pass. If an API workload is cost-sensitive, developers may reduce context size and retrieval breadth. Those decisions are valid, but they must be explicit. A cheaper answer with weaker source coverage should not be labelled as fully verified.

The practical recommendation is to write evidence status into the answer. Phrases such as “official pricing confirms”, “public documentation does not specify”, “the cited source supports only part of this claim” and “manual verification required” make uncertainty visible. That is better than hiding uncertainty behind a neat citation list.

Takeaways

• Use “cite every factual claim” instead of vague requests such as “add sources”.
• Separate source collection, claim extraction and final drafting for any high-risk topic.
• Open citations manually because a clickable link does not prove sentence-level support.
• Treat API citation handling as a parser and rendering problem, not only a prompt problem.
• Budget for verification queries because source-first workflows consume more Pro, Research or API capacity.
• Verify nouns and numbers first: prices, names, dates, plan caps, quotes, laws and statistics.
• Label uncertainty directly when Perplexity cannot confirm a claim from an authoritative source.
• Use Perplexity as an evidence triage assistant, not the final authority for publishable claims.

Our Content Testing Methodology

During our 2026 evaluation, this guide was compiled by replicating citation-failure scenarios across Perplexity web prompts, source-first prompting patterns and API-style response handling described in Perplexity documentation. We cross-checked official plan limits, enterprise pricing, API pricing, media constraints and citation parsing guidance against current Perplexity documentation, then compared those documented behaviours with public developer issue reports where citation objects were missing, mis-rendered or mis-mapped in third-party clients. Benchmark and risk claims were checked against generative search research, the EBU and BBC news integrity study, Stanford HAI’s AI Index and named 2025-2026 reporting. The workflow recommendations were included only where the test condition could be reproduced as a clear operational step: source collection, claim extraction, inline citation drafting, manual page inspection, API parser validation or audit logging.

Conclusion

Perplexity not citing sources is best understood as a signal, not a single bug. Sometimes the prompt did not demand sentence-level support. Sometimes the selected mode produced a broad synthesis. Sometimes the source page could not be parsed cleanly. Sometimes the API returned citation data that a client failed to display. The user’s job is to diagnose which layer broke, then tighten the workflow rather than simply regenerate and hope.

The future of AI search will not be decided by whether every answer shows links. It will be decided by whether users, editors and developers can trace claims back to evidence reliably enough for the task at hand. Perplexity remains valuable because it compresses source discovery and turns research into a conversational flow. Its limitation is that conversation can make evidence feel settled before it has been checked. The balanced approach is neither blind trust nor rejection. It is source-first use, claim-level verification and explicit uncertainty when the evidence does not support the prose. Open questions remain around citation mapping in complex API workflows, access to paywalled sources, source freshness and how AI systems should represent conflicting evidence. Those are product and governance questions, not just prompt-writing problems.

FAQs

Why Is Perplexity Not Citing Sources?

Perplexity may omit visible citations when the prompt is broad, the answer synthesises general background, the selected mode retrieves only a small source set, or the interface fails to render citation objects. Ask for inline citations after every factual sentence and use a source-first pass for important work.

How Do I Force Perplexity to Cite Sources?

Write the citation requirement in the first line of the prompt. Use: “Use only verifiable sources, place a citation after every factual claim and say when evidence is missing.” For complex work, ask for sources only before asking for the final answer.

Can Perplexity Citations Be Wrong?

Yes. A citation can be topically relevant but fail to support the exact sentence. It can also point to outdated, weak or mis-mapped evidence. Always open the cited page and compare the claim with the source text before relying on it.

Does Pro Search Improve Citation Accuracy?

Pro Search can improve retrieval depth and source discovery, but it does not remove the need for manual verification. Better source coverage helps only when the answer is checked at claim level and the cited pages support the exact statements.

Why Are Perplexity API Citations Missing in My App?

The API may return citation or search result fields that a third-party client does not parse or display. Inspect the raw response, parse citation objects separately and map every footnote marker to a stored source URL.

Can Perplexity Cite a Specific Page in a PDF?

It can often summarise uploaded files, but page-specific citation quality depends on the file format, extraction quality and interface. For high-stakes work, ask for page, section and phrase evidence, then manually inspect the PDF.

Should I Cite Perplexity or the Original Source?

For published research, cite the original source whenever Perplexity helped you discover it. Cite Perplexity itself only when the AI-generated interaction is part of your method or evidence record and your style guide allows it.

What Should I Do When a Citation Is Missing?

Ask which source supports the exact sentence, regenerate with a narrower source-first prompt, or remove the unsupported claim. If using the API, check whether citation fields exist in the raw payload but are not rendered by the client.

References

Chatbox AI. (2025, June 20). [BUG] Perplexity citations not shown. GitHub. https://github.com/chatboxai/chatbox/issues/2404

European Broadcasting Union. (2025). News integrity in AI assistants: An international PSM study. https://www.ebu.ch/research/open/report/news-integrity-in-ai-assistants

Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating verifiability in generative search engines. arXiv. https://arxiv.org/abs/2304.09848

Malik, A. (2025, June 5). Perplexity received 780 million queries last month, CEO says. TechCrunch. https://techcrunch.com/2025/06/05/perplexity-received-780-million-queries-last-month-ceo-says/

Perplexity AI. (2026). Enterprise pricing. https://www.perplexity.ai/enterprise/pricing

Perplexity AI. (2026). Pricing. Perplexity API docs. https://docs.perplexity.ai/docs/getting-started/pricing

Perplexity AI. (2026). Streaming citation parsing. Perplexity API docs. https://docs.perplexity.ai/docs/cookbook/articles/streaming-citations/README

Perplexity AI Help Center. (2026). Which Perplexity subscription plan is right for you? https://www.perplexity.ai/help-center/en/articles/11187416-which-perplexity-subscription-plan-is-right-for-you

Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index report. https://hai.stanford.edu/ai-index/2026-ai-index-report