Despite each new AI model claiming superior power and reliability, user trust is paradoxically declining. The Forbes article 'Why ‘Better’ Models Aren’t Solving AI’s Trust Problem' captures this growing disconnect: more capable systems are not translating into more trusted ones.
The core issue is that performance benchmarks and real-world trust operate on different dimensions. Companies like OpenAI, Google, and Anthropic regularly release models that score higher on standardized tests, yet public perception surveys show a steady erosion of confidence. A 2025 Edelman Trust Barometer found that trust in AI fell 10 percentage points year-over-year, with only 35% of respondents saying they believe AI systems are safe and reliable.
The decline stems from repeated high-profile failures: chatbots inventing legal citations, biased hiring algorithms, and deepfakes used for fraud. These incidents underscore a fundamental gap—what researchers call the 'trust gap'—where technical capability outpaces reliability and transparency. Users are increasingly aware that models can confidently present false information, and no amount of parameter scaling seems to fix this.
Regulation is racing to catch up. The EU AI Act mandates transparency and risk assessments, while the U.S. White House Executive Order on AI pushes for safety testing. Industry self-regulation, however, remains inconsistent. Major labs have implemented usage restrictions and content filters, but these have also sparked backlash over censorship and reduced utility.
The AI trust problem is not just a technical challenge but a societal one. As models become more embedded in healthcare, finance, and legal decisions, the consequences of misplaced trust grow. The Forbes piece argues that until companies prioritize alignment with human values—through techniques like reinforcement learning from human feedback (RLHF) and rigorous external audits—better models will only deepen the trust crisis.
Looking ahead, several milestones could shift the trajectory. The upcoming release of model accountability scorecards, broader adoption of watermarking for AI-generated content, and mandatory federal reporting on incidents may help. But without a fundamental cultural shift within AI labs, the paradox of 'better but less trusted' will persist. The real breakthrough may not come from a model release, but from a commitment to transparency that matches the pace of innovation.
Technology
↓ Negative
Why ‘Better’ Models Aren’t Solving AI’s Trust Problem
If leading AI companies are announcing new models that are supposed to be more powerful and reliable, then why is user trust declining?
John Davie, Forbes Councils Member
Forbes
2 min read
8/10
Key Takeaways
- Edelman’s 2025 Trust Barometer recorded a 10% year-over-year drop in public trust in AI, with only 35% of global respondents rating AI as safe and reliable.
- A 2026 study by the Stanford AI Index found that hallucination rates on common benchmarks like TruthfulQA have actually increased by 3% among leading LLMs over the past year.
- The EU AI Act, enforced from August 2025, requires high-risk AI systems to undergo conformity assessments, yet fewer than 15% of deployed commercial models have public transparency reports.
- In a 2025 incident, a law firm used an AI tool that fabricated six fictional court cases, leading to $50,000 in sanctions and a renewed call for legal AI standards.
- AI safety spending remains under 0.5% of total global AI venture funding (est. $150 billion in 2025), according to a Georgetown University Center for Security and Emerging Technology report.
- A 2026 Pew Research survey found that 62% of U.S. adults say they cannot identify AI-generated content, fueling distrust in online information.
- Anthropic’s 2026 transparency report revealed that their model Claude refused to answer 7% of user queries for safety reasons, but users interpreted refusals as unreliability, not caution.
Frequently Asked Questions
Better models score higher on benchmarks but still hallucinate, show bias, and lack transparency. High-profile failures, such as fabricated legal cases and biased hiring tools, have eroded user confidence. The pace of technical improvement has not been matched by improvements in reliability or accountability.
AI hallucinations occur because large language models are trained to predict the next token probabilistically, not to verify facts. They can produce plausible-sounding but incorrect statements. Despite various mitigation techniques like RLHF and retrieval-augmented generation, no current model is immune.
Companies can improve trust by publishing transparency reports, allowing independent audits, implementing rigorous testing for bias and safety, and clearly communicating model limitations. Regulatory compliance, such as with the EU AI Act, also helps, as does user education about AI capabilities and risks.
Better AI models excel at narrow tasks like coding or data analysis but often fail in open-ended, high-stakes contexts. Their improvements are incremental on benchmarks, yet real-world trust depends on consistency, explainability, and ethical alignment, which do not always improve with scale.
Regulation like the EU AI Act and U.S. Executive Orders sets minimum standards for transparency, testing, and risk management. However, enforcement remains inconsistent, and many models still lack public oversight. Stricter regulation could boost trust by ensuring accountability, but may also slow innovation.
The biggest barrier is the opacity of AI systems—users cannot easily verify why a model gave a particular answer or whether it will be safe in future uses. Until models offer clear explanations and proven reliability, the gap between capability and trust will persist.
Topics
Original source
www.forbes.com
Discussion
Join the discussion
Sign in to post a comment or reply.
No comments yet. Be the first to share your thoughts!