Trump Officials Deploy AI to Slash Federal Regulations

The Trump administration, through the Department of Government Efficiency (DOGE), deployed an AI pipeline to identify, prioritize, and draft deregulatory actions across the federal code. Internal slides and newly released records describe a DOGE AI Deregulation Decision Tool and a related system labeled SweetREX intended to review roughly 200,000 regulatory items and eliminate 50% deemed not statutorily required. The materials claim the system can process more than 100,000 public comments in under half an hour and cut human review time by 93%. Democracy Forward and agency staff reported failures where the AI misread statutory language and produced legally questionable outputs.

The documents do not publish full model architecture or training data, but they imply the use of large language models and rule-based classifiers fine tuned to policy objectives. SweetREX is described as programmed to prioritize rules that 1) impose costs on private enterprise, 2) limit business innovation, or 3) use race-based classifications. Reported capabilities include rapid text classification of regulatory sections, automated summarization of rule histories, bulk synthesis of public comments, and draft rule language generation. Practitioners should note three technical risk areas:

• •Model hallucination and misinterpretation of statutory text when legal precision is required
• •Dataset and label bias if training data reflect a deregulatory objective rather than neutral legal standards
• •Lack of audit trails and provenance for generated draft rules, which complicates judicial review

This deployment is a precedent for direct automation of administrative law decisions. The administrative process under the Administrative Procedure Act requires record building, reasoned explanation, and responsiveness to comments; automated summaries or draft rescissions cannot substitute for the legal standards courts apply. Using AI to scale deregulatory throughput changes the workflow incentives of agencies; it shifts the bottleneck from research and rulemaking deliberation to legal defense and litigation. The plan also echoes wider trends where government offices adopt narrow objective functions for models, sometimes without commensurate guardrails. Legal and agency staff already flagged errors, highlighting how aggressively optimized systems can produce high-impact false positives in regulatory removal.

For ML engineers and policy technologists, this case crystallizes governance requirements when models touch law and rights. You need documented training corpora, red-team evaluations for legal reasoning, metrics beyond accuracy such as false-positive rate for statutory misread, deterministic explanations for each recommendation, and human-in-the-loop controls that can be audited. The practical stakes are litigation risk, loss of public protections, and potential invalidation of agency actions if courts find the record inadequate.

Expect FOIA releases, congressional oversight, and litigation testing the admissibility of AI-assisted rulemaking. Monitor whether agencies publish model specs, evaluation benchmarks, or commit to open audits. Watch for administrative guidance from OMB or DOJ on permissible automated assistance in rulemaking, and for courts to define evidentiary standards for AI-generated administrative records.