Perly Consulting │ Beck Eco

The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY

The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.

The Daily Dispatch

A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.

AI Maturity by Domain

Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail

DOMAIN
BLEEDING EDGEESTABLISHED

Hallucination detection & factuality assessment

LEADING EDGE

TRAJECTORY

Stalled

Tools and processes for detecting AI-generated hallucinations and assessing the factual accuracy of model outputs. Includes automated fact-grounding and source verification; distinct from fact-checking in research which verifies human-authored rather than AI-generated claims.

OVERVIEW

Hallucination detection has reached the point where forward-leaning organisations are deploying it in production, but most have not started — and the gap between vendor claims and real-world reliability remains the practice's defining tension. Detection and factuality assessment techniques aim to identify when LLMs generate plausible-sounding false claims. GA tooling now exists from AWS, Vectara, Datadog, and Microsoft, yet independent research consistently shows that benchmark accuracy does not predict production performance: benchmarks use artificially short prompts and simple contexts, while real systems operate at two to five times that complexity. The result is a field where platform maturity has outpaced measurement credibility. No single detection method generalises across languages, domains, and reasoning types, and leading models still hallucinate at 15-16% baselines on realistic evaluations. Detection is a necessary governance control, but it functions as one layer in a stack that still requires human oversight and external grounding — not a standalone solution.

CURRENT LANDSCAPE

The vanguard of adoption centres on a handful of named deployments. DoorDash fields hundreds of thousands of AI-enabled customer calls daily using RAG with hallucination detection; AWS Bedrock trials block 75% of unsupported answers; and Vectara's HHEM model has surpassed 100,000 downloads, offering real-time factual consistency scoring. Enterprise adopters are beginning to integrate hallucination risk scores as formal QA gates in production pipelines, with early reports of 40% fewer escalations. Real-world implementations like Qdrant's three-layer defense demonstrate 94% hallucination reduction across five enterprise clients in production. These are meaningful signals, but they represent the frontier — not the field.

The research frontier has expanded beyond text. Amazon Science's VADE detects hallucinations in vision-language models via attention maps; ACL 2026 introduced HalluAudio, the first large-scale audio hallucination benchmark spanning speech, environmental sound, and music with 5K+ human-verified pairs. Financial long-context detection advanced with NeurIPS 2025's PHANTOM benchmark, testing hallucination detection on 500-30K token SEC filings—revealing that small models still score near-random and Lost-in-the-Middle degradation remains unresolved at scale. Amazon Science also published real-time detection methods for speech LLMs using attention-derived metrics, extending detection to the spoken-word domain. Multimodal expansion signals the field's coming-of-age, but each modality surfaces new measurement challenges.

The credibility problem runs deeper than any single vendor. AWS claims 99% verification accuracy for Bedrock Automated Reasoning; Vectara reports sub-1% hallucination rates for small models. Independent evaluations tell a different story: frontier models expose a benchmark-accuracy paradox—GPT-5.5 tops every benchmark yet hallucinate at 86%, a critical negative signal that accuracy-only metrics structurally reward confident guessing over calibrated uncertainty. Cross-lingual evaluations show hallucination rates 15-35% higher in non-English languages, exposing that detection methods trained on English corpora fail systematically in low-resource language contexts. EACL 2026 peer-reviewed research (April) confirms that detection techniques measure consistency (with 50%+ inconsistency in benchmarks like Med-HALT), not correctness—fundamentally misaligned with what practitioners need. The AuthenHallu benchmark reveals only 60% detection accuracy even on SOTA models against authentic LLM-human interactions. Detection also degrades where enterprises need it most: grounding evaluation breaks down in multi-turn agentic workflows (HalluHard benchmark shows 30%+ hallucination even with web search), and no method yet handles the intersection of multilingual, multi-domain, and extended-context inputs. A stark real-world signal: ICLR 2026 integrity crisis documented 50+ peer-reviewed papers accepted to a top venue despite containing hallucinated citations and fabricated datasets—a failure of detection governance at the research infrastructure level. Detection methods optimize for detectable consistency while missing correctness gaps, making deployment without mandatory human oversight unsafe in high-stakes domains. The field has reached an inflection: detection is GA, widely available, and necessary—but fundamentally insufficient without human oversight and governance layers.

TIER HISTORY

ResearchJan-2023 → Jul-2023
Bleeding EdgeJul-2023 → Feb-2026
Leading EdgeFeb-2026 → present

EVIDENCE (90)

— Amazon Science peer-reviewed inference-time hallucination detection for speech LLMs using attention-derived metrics (AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY), advancing real-time detection in spoken-word contexts.

— Critical assessment showing frontier model (GPT-5.5) achieves highest accuracy but 86% hallucination rate, with Nature paper evidence that accuracy-only benchmarks structurally reward confident guessing over calibrated uncertainty—negative signal on detection sufficiency.

— Practitioner analysis documenting production discrepancies: GPT-5.5 achieves 57% accuracy but 86% hallucination rate on same benchmark, showing accuracy and hallucination as separate dimensions requiring independent assessment.

— NeurIPS 2025 Datasets & Benchmarks benchmark for hallucination detection on 5K+ SEC filing QA pairs at 500-30K token lengths, revealing Lost-in-the-Middle degradation and that small models score near-random—addressing critical gap in long-context financial hallucination detection.

— ACL 2026 first large-scale audio hallucination benchmark (5K+ human-verified QA pairs) across speech, environmental sound, and music; systematic evaluation demonstrating that hallucination detection research now spans text, vision, and audio modalities.

— Practitioner analysis documenting hallucination rates 15-35% higher in non-English languages (38-point deficits in low-resource languages), exposing evaluation gaps in multilingual production systems.

— Real-world detection failure at scale: 50+ ICLR papers contained hallucinated citations and fabricated datasets despite peer review, documenting governance lessons and detection gaps applicable to enterprise systems.

— Amazon Science research presenting attention-map-based hallucination detection for Vision Language Models by identifying misalignment between outputs and visual content, extending detection methods to multimodal domain.

HISTORY

  • 2023-H1: Hallucination detection field established with competing research methodologies (SAC3, SRLScore, HalluMix benchmark). Critical analyses exposed failures in QA-based approaches and existing benchmarks; multilingual and long-context detection gaps identified. Commercial products (Vectara, Bedrock AI) launched with retrieval-augmented generation as primary mitigation. ChatGPT's systematic factuality failures documented in production use.
  • 2023-H2: Research methodologies refined with uncertainty-based, early-detection, and cognitive approaches achieving high performance metrics. Domain-specific benchmarks (DelucionQA) emerged for RAG scenarios. Vectara HHEM v2 reached 40k+ monthly downloads, indicating significant product adoption. Community tracking expanded with curated research resources (1k+ stars). Vendors began packaging detection into platform features (AWS Guardrails preview). Fundamental barriers to adoption persisted: no general-purpose cross-domain, multilingual detection solution; RAG-based mitigation remained dominant strategy.
  • 2024-Q1: AWS Bedrock Guardrails GA with contextual grounding checks entered the market, signaling mainstream platform adoption. Benchmarks quantified capability gaps: HallusionBench showed GPT-4V at 31.42% accuracy on multimodal detection. Global-Liar documented deployment risks including time-based performance regression and geographic bias. Real-world adoption metrics (Applause 38%, Aporia 89%) confirmed hallucinations as a widespread operational blocker. Research advanced assessment methodologies with peer-reviewed LLM-based fact-checking studies, though inconsistent accuracy across claim types. Platform-driven RAG integration and research-stage detection methods continued in parallel tracks with no convergence toward a general-purpose solution.
  • 2024-Q2: Independent empirical evaluations revealed critical limitations of enterprise deployment. Stanford-Yale study found legal AI tools (Lexis+, Westlaw) hallucinating at 17-33% rates despite vendor reliability claims. Medical literature study documented ChatGPT/Bard at 39.6-91.4% hallucination rates in systematic reviews, with researchers concluding LLMs unsuitable as primary tools. New detection methods advanced: Oxford semantic entropy probes reduced computational cost. Enterprise confidence metrics declined further (68% of data professionals lack data quality assurance). No convergence toward general-purpose detection solution; field remained stratified between platform-integrated RAG and research-stage methods.
  • 2024-Q3: Platform vendors accelerated product maturity: AWS Bedrock Guardrails announced expanded detection (July); Vectara released HHEM-2.1 claiming performance gains. Research communities published multiple detection methods at ACL 2024 (zero-resource, unsupervised real-time approaches) and Nature Machine Intelligence survey elevated field discourse. Critical discovery: leading models (Claude-3.7, GPT-o1) demonstrated only 81-82% reasoning factual accuracy, revealing hallucination as a fundamental model architecture constraint rather than a detection-solvable problem. No progress toward general-purpose solution; field remained fragmented across platform integrations, open-source tools, and research prototypes.
  • 2024-Q4: Vendors intensified product development: AWS introduced automated reasoning checks in Bedrock Guardrails (December re:Invent) claiming 85% block rate; RELAI launched commercial hallucination detection agents. Research synthesis accelerated: comprehensive surveys (arXiv, EMNLP multimodal) synthesized field taxonomy; cost-effectiveness analysis emphasized performance-budget trade-offs; NeurIPS papers (LLM-Check, HaloScope) advanced internal-representation and self-supervised detection methods. Deployment challenges remained unchanged: enterprise legal AI hallucinations persistent at 17-33%; multimodal detection accuracy still below 35%. Consensus emerged: no single detection method universal; layered approaches (detection + grounding + oversight) necessary. Market consolidation visible: major cloud platforms embedded detection as native capability; research-to-product lag persisted at 12-18 months; general-purpose cross-domain solution remained absent.
  • 2025-Q1: Platform vendors embedded detection deeper: AWS Bedrock launched RAG Evaluation with hallucination detection (faithfulness) as core metric (March); Vectara upgraded HHEM factual consistency scoring with claims of 100k+ downloads. Vendor ecosystem expanded: Cisco Research released open-source PolygraphLLM toolkit citing Air Canada penalty and 3-10% critical-domain hallucination rates. Research revealed domain-specific detection gaps: HalluCounter achieved >90% average confidence (March) but SelfCheck-Eval discovered methods fail on mathematical reasoning, introducing AIME Math Hallucination benchmark (February). Critical adoption finding: Gartner predicted 30% GenAI project abandonment by year-end, citing inadequate risk controls—hallucination detection surfaced as key blocker despite platform product maturity. Field consensus shifted: not "which detection method" but "detection as necessary but insufficient component of layered approach."
  • 2025-Q2: Vendor product expansion masked a measurement crisis: Datadog launched LLM Observability with hallucination detection (May); Vectara released Hallucination Corrector with claimed 0.9% hallucination rates (May); HHEM reached 250k+ downloads. However, EMNLP 2025 peer-reviewed research (April-June) revealed that hallucination detection metrics themselves fail to align with human judgments across 37 models—undermining confidence in detection system evaluation. Multilingual study of 61,514 claims (June) exposed critical vulnerability: GPT-4o declined 43% of claims and misclassified factual content more than opinions, revealing LLM-based fact-checking as fundamentally unreliable. Real-world incidents intensified: Air Canada tribunal ruling, DPD/Virgin Money chatbot failures, Cursor policy hallucinations (May) documented persistent deployment failures. Pacific Northwest National Laboratory case study (June) showed Bedrock Knowledge Bases achieving only 0.3% precision on basic retrieval until switching to synthetic data. Field sentiment: hallucination detection shifted from engineering problem to persistent architectural constraint requiring governance, not just technical innovation.
  • 2025-Q3: ACL 2025 research (July) exposed fundamental flaws in detection evaluation: state-of-the-art factuality metrics are inconsistent, misestimate accuracy, and exhibit biases against paraphrased outputs. FactBench dynamic benchmark demonstrated scale does not guarantee factuality (Llama-3.1-405B underperformed 70B variant); meta-analysis revealed ROUGE-based evaluation is misleading with reported progress gains potentially illusory. AWS Bedrock Automated Reasoning GA (August) claimed 99% verification accuracy but practitioner testing revealed inconsistency. Microsoft VeriTrail methodology (accepted ICLR 2026) advanced detection in multi-step workflows. Vectara leaderboard credibility eroded: HHEM-2.1-Open self-reported F1 only 45-66%. Legal services case study showed AWS Guardrails grounding evaluation degrades in multi-turn agentic RAG. Market data showed $765M market in 2024, forecast $6.2B by 2033, but adoption bottlenecked by evaluation methodology instability. Consensus solidified: platform features mature but scientific foundations for validating detection reliability unstable.
  • 2025-Q4: Research methods diversified but gap between benchmarks and production persisted. Academic publications (HalluCounter achieving >90% accuracy in ACL Findings, Cambridge Consortium multi-LLM ensemble approaches, lightweight HALT probes with <0.1% overhead) showed continued methodological innovation, but independent critical analyses exposed the core problem: hallucination benchmarks (RAGTruth, FaithBench) use unrealistically simple document contexts with median prompt lengths under 550 tokens compared to 2000-3000+ token production systems, making benchmark performance unreliable predictors of real deployment success. Persistent baseline hallucination rates (GPT-4o ~15.8%, Claude 3.7 ~16% on real benchmarks) indicated detection had not addressed the fundamental LLM unreliability. Enterprise adoption signals mixed: Saison-Vectara partnership for conversational AI indicated vendor willingness to co-deploy, but existing case studies showed multi-turn agentic workflows remained problematic as grounding evaluation degraded with conversation length. Vendor product maturity continued (Vectara Hallucination Corrector, AWS Bedrock Automated Reasoning GA) with widespread availability but claims remained bounded by unclosed gap between benchmarks and production. Market forecasts unchanged ($6.2B by 2033) but adoption remained constrained by recognition that detection cannot fully substitute for model-level reliability, requiring mandatory human oversight in high-stakes domains. By year-end 2025, the field had hardened into uncomfortable stability: detection is GA, widely available, and architecturally mature—but benchmark claims are not reliable predictors of production performance and the fundamental architectural problem (LLM hallucination as model-level constraint) remains unsolved.
  • 2026-Jan: Vendor product maturation accelerated with real-world adoption signals. Vectara launched Factual Consistency Score powered by HHEM (100,000+ downloads) offering real-time 0-1 scoring; DoorDash deployed enterprise-wide contact center AI fielding 100,000s of daily customer calls using Claude 3 Haiku with RAG and detection; enterprise adoption patterns showed hallucination risk scoring now used as formal QA gates in production pipelines (AWS Bedrock trials blocking 75%, early adopters reducing escalations 40%). Research advanced detection methodologies: PEFT fine-tuning consistently strengthened detection across models; KnowHalu multi-phase framework achieved 82.2% accuracy. However, credibility gaps widened: user surveys (Duke, 94% report accuracy varies significantly) exposed disconnect between vendor claims and real-world experience. Critical assessment surfaced persistent fundamental issues—benchmark incentives reward guessing, training data contradictions, pragmatic failures—indicating detection remains necessary but insufficient control requiring mandatory human oversight. Field consensus held: platform maturity established but benchmark-to-production gap unresolved, adoption constrained by recognition that detection cannot substitute for model-level reliability.
  • 2026-Feb: Vendor product maturation accelerated with deepening deployment signals. Vectara launched Factual Consistency Score (100k+ downloads, 0-1 real-time scoring); DoorDash deployed enterprise-wide contact center AI fielding 100,000s daily calls using Claude 3 Haiku with RAG and detection. Research methodologies continued advancing: HALT lightweight detector achieved 60x speedup; VIGIL introduced fine-grained multimodal detection benchmark. However, credibility gaps widened decisively: Open Data Institute study testing 22,000+ prompts found models providing false answers, rarely admitting uncertainty; critical analyses exposed vendor claim misalignment (o3 at 33-51% error rates vs. 99% accuracy claims; Mata v. Avianca case; enterprise failures in legal, medical, financial domains). Field consensus hardened: detection is GA and widely deployed but benchmark-to-production gap remains unresolved; detection cannot substitute for model-level reliability. Early 2026 marked inflection point—platform maturity established but user experience data revealed persistent fundamental limitations requiring mandatory human oversight.
  • 2026-Apr: Market growth signals ($1.86B to $2.47B at 33.2% CAGR) and production architectures advanced, with Qdrant three-layer enterprise deployments demonstrating 94% hallucination reduction and Amazon Science publishing FINCH-ZK cross-model consistency detection improving F1 scores by 6–39%, plus new inference-time detection for speech LLMs via attention-derived metrics (AUDIORATIO, AUDIOCONSISTENCY, AUDIOENTROPY). A critical benchmark-accuracy paradox sharpened: GPT-5.5 topped every benchmark yet hallucinated at 86%, with Nature paper evidence that accuracy-only benchmarks structurally reward confident guessing — confirming accuracy and hallucination rate are independent dimensions. The Charlotin incident database now tracks 1,200+ real-world hallucination incidents (5–6 new entries daily), EACL 2026 research confirmed detection methods measure consistency not correctness, the HalluHard benchmark showed frontier models hallucinate 30%+ even with web search, and OpenAI research characterised hallucinations as mathematically inevitable under current architectures — reinforcing that detection remains a necessary governance layer but not a solution to the underlying model reliability problem.

TOOLS