TDTI Research Series

Independent research on trusted data and AI.

Our papers test public claims against public data: what governments and firms disclose about AI, how AI agents can be governed, and what workforce evidence shows. Each summary below links to the full paper.

AI governance & public accountability

What governments and firms disclose about AI

Independent tests of public AI disclosures against other public records: contracts, incident reports, other jurisdictions' inventories and consumer complaints.

RESEARCH PAPER

August 2026

Joseph Atalla, PhD

12 pages · PDF

Same Tool, Different Verdict?

A reproducible audit of risk classification and disclosure across U.S. public-sector AI inventories

Does the same AI product receive the same risk treatment in different governments? This audit compares five reporting mechanisms (federal, New York State, New York City, California and San José), using the complete 19-system New York State inventory as a fixed starting point.

Reuse of the same products is real: Microsoft Copilot and Google Translate appear in all three current federal, state and city inventories. But only two of the five mechanisms publish an explicit risk or high-impact field, and the federal PATTERN tool carries opposite high-impact ratings in two Justice Department offices. The paper concludes that the main barrier to AI transparency is the lack of shared identifiers, risk fields and evaluation evidence, not the lack of inventories.

4 of 19New York State systems with a verified exact product match in the NYC or federal inventory

2 of 5reporting mechanisms with an explicit risk or high-impact field

1federal tool rated both high-impact and not high-impact in 2025

Read the full paper PDF · 12 pages · 1.0 MB

RESEARCH PAPER

October 2026

Joseph Atalla, PhD

9 pages · PDF

Following the Money

Can federal AI disclosures be traced to federal contracts?

Agencies report their AI use cases to OMB and their contracts to the Federal Procurement Data System, but nothing connects the two. This audit classifies every contract identifier in the 2024 and 2025 federal AI inventories and verifies a random sample of 60 identifiers against the procurement record.

Only about half of contractor-built AI use cases listed a usable identifier, an estimated 28% can be traced to a specific contract action, and few contracts describe AI at all. The riskiest systems were the least traceable, and the 2025 inventory dropped the identifier question entirely. The paper proposes a five-tier Contract Traceability measure and six low-cost fixes.

15%of 351 rights- or safety-impacting use cases (2024) listed a usable contract identifier

≈28%of contractor-built AI use cases trace to a specific contract action

7%of verified contracts mention AI, machine learning or algorithms

Read the full paper PDF · 9 pages · 937 KB

RESEARCH PAPER

October 2026

Joseph Atalla, PhD

9 pages · PDF

Incidents Without Inventories

Do government AI failures appear in public AI disclosures?

If disclosure works, the AI systems that harm the public should already appear in government AI inventories. This study identifies the 113 incidents in the AI Incident Database where a U.S. government body deployed the system and checks each federal system against the 2023, 2024 and 2025 federal inventories.

Six of 13 civilian federal systems with incidents are listed by name, but four never appear. Most incidents involve local police, sheriffs and school districts, and 85% of state, local and court incidents involve bodies with no duty to publish an AI inventory. Eleven federal incidents since 2025 come from staff use of general-purpose generative AI or AI-made images in official work, which inventories are not designed to capture. The paper proposes six fixes that connect incident reporting to disclosure.

113AI incidents where a U.S. government body deployed the system

4 of 13civilian federal systems with incidents never appear in a federal inventory

85%of state, local and court incidents involve a body with no inventory duty

Read the full paper PDF · 9 pages · 1.0 MB

RESEARCH PAPER

September 2026

Joseph Atalla, PhD

9 pages · PDF

Complaints About the Machine

What consumers told the CFPB about financial firms' chatbots, phone systems and automated decisions, 2019–2026

This study screens all 3.49 million complaint narratives the Consumer Financial Protection Bureau published for complaints received from January 2019 to August 2026, and validates the screen against a hand-coded sample.

Between 1% and 5% of checking, savings and card complaints describe a firm's automated system as part of the problem. Phone systems that block a person and automated decisions such as fraud freezes account for most cases, chatbots for about 10%. Half of coded cases say the consumer could not reach a human, and a quarter describe losing access to their own funds. Because the CFPB stopped publishing narratives on August 14, 2026, the study cannot be repeated with the public data that remain.

1%–5%of bank and card complaint narratives describe a chatbot, phone system or automated decision

52%of coded cases say the consumer could not reach a human

45% → 0share of complaints published with a narrative, 2019 to August 2026

Read the full paper PDF · 9 pages · 878 KB

Trusted AI & agentic systems

Keeping AI agents reliable and governed

A system-level approach to evaluating, observing and governing AI agents that act through tools, tested in a controlled experiment.

RESEARCH PAPER

October 2026

Joseph Atalla, PhD

9 pages · PDF

Can a Reliability Control Plane Make Agentic AI Safer Without Making It Less Effective?

A controlled fault-injection study with external governance-benchmark triangulation

Do runtime safety controls make AI agents less useful? This paired experiment runs 10,000 episodes across 20 task types with identical injected faults, once without runtime governance and once with RCP-Lite's policy interception, approval gates, process checks, budget controls and rollback.

Executed policy and process violations fell to zero, budget compliance rose to 99.88%, and task completion was not worse (93.8% without controls, 94.5% with them). The fault rates are experimental settings, not measurements of how often any production model fails, so the effect sizes need replication with live models.

10,000paired episodes across 20 task types

+0.72 ppchange in task completion with controls (95% CI 0.33 to 1.11)

0.00%executed policy and process violations with RCP-Lite

Read the full paper PDF · 9 pages · 702 KB

RESEARCH PAPER

October 2026

Joseph Atalla, PhD

12 pages · PDF

Reliability Control Plane for Agentic AI

Integrating evaluation, observability and governance for long-horizon LLM agents

AI agents that plan, call tools and act over many steps fail in ways conventional testing and monitoring miss. This paper reviews the three pillars of agent reliability (evaluation, observability and governance) and proposes a Reliability Control Plane that joins trace-based evaluation, standardized telemetry and runtime policy enforcement into one continuous assurance loop, drawing on the NIST AI RMF and its Generative AI Profile.

A lightweight reference implementation, RCP-Lite, raised task success from 94.2% to 100% in a 1,000-episode proxy evaluation and stopped every executed policy, process and budget violation. The paper also proposes Continuous Sentinel Evaluation, a canary-testing pattern for agents already in production.

1,000balanced proxy episodes across 20 tasks

+5.8 pptask success with RCP-Lite (94.2% to 100%)

0executed policy, process or budget violations with RCP-Lite

Read the full paper PDF · 12 pages · 767 KB

Organizational systems & future of work

People, support and the capacity to change

Quantitative research on the federal workforce: what predicts employee outcomes under telework, and whether workforce climate explains AI adoption.

RESEARCH PAPER

September 2026

Joseph Atalla, PhD

9 pages · PDF

Support-Integrated Telework in Public Agencies

A quantitative analysis of the 2023 Federal Employee Viewpoint Survey and the Support-Integrated Telework Model (SITM)

Telework debates usually count days in the office. Using survey-weighted models of 425,815 federal employees in the 2023 Federal Employee Viewpoint Survey, this paper finds that perceived organizational support is the largest and most consistent correlate of engagement, job satisfaction, performance confidence and perceived innovation, while telework frequency shows much smaller, outcome-specific associations.

It introduces the Support-Integrated Telework Model (SITM), which treats telework as one work-design condition inside a broader support environment. The formal research edition of 2026 doctoral research; findings are associational, not causal.

425,815federal employee records analyzed

4employee outcomes modeled

.605–.793standardized organizational-support coefficients across outcomes

Read the full paper PDF · 9 pages · 563 KB

RESEARCH PAPER

August 2026

Joseph Atalla, PhD

12 pages · PDF

Does AI Capacity Require Human Capacity?

Federal AI adoption and workforce climate across U.S. agencies

Do agencies with more engaged workforces put more of their AI into use? This exploratory study links 2024 Federal Employee Viewpoint Survey engagement scores to the 2025 federal AI use case inventory for 11 agencies and 1,810 use cases.

Engagement showed no reliable association with the share of AI use cases deployed or piloted. That does not mean people are unimportant to AI; it means broad workforce climate is a poor stand-in for AI readiness, which needs direct measures of technical talent, training, governance, data, infrastructure and acquisition capability.

11agencies in the primary sample

1,810individually reported AI use cases

−0.09correlation between engagement and deployment share (p = 0.79)

Read the full paper PDF · 12 pages · 960 KB

RESEARCH PAPER

September 2026

Joseph Atalla, PhD

8 pages · PDF

Telework and Organizational Support

A leadership agenda for large institutions

The leadership edition of the telework research. Across four employee outcomes, perceived organizational support was the strongest and most consistent predictor; telework frequency had smaller associations and no significant relationship to engagement once support was accounted for.

The case study turns those findings into a proposed operating model for large institutions: set work arrangements by mission and role, and manage support quality (training, resources, supervisor work-life support, senior-leader endorsement and management communication) as a measurable responsibility, with a roadmap and scorecard for local validation.

425,815federal employees analyzed

5dimensions of organizational support made operational

4 of 4outcome models where support was the strongest predictor

Read the full paper PDF · 8 pages · 694 KB

About these papers. TDTI research is independent and analyzes publicly available data unless a paper states otherwise. The summaries on this page are condensed from each paper; the PDF is the complete and authoritative version, including methods, limitations and sources. How TDTI is governed and funded, and how to report an error, is on our Governance & independence page.

Work with the Institute

Discuss our research or a collaboration.

Contact the Institute