Does AI still hallucinate or is it becoming more reliable? Pg20

1. What are AI Hallucinations?

• Hallucinations occur when AI generates incorrect, made-up, or factually wrong content with high confidence.
• Examples: ChatGPT models generating fake pizza tips or creating “rooms with no elephants” that still include elephants in some form.

2. Why Does it Happen?

• Hallucinations stem from inadequate training data, especially for rare or ambiguous queries.
• Models try to guess answers even when they lack relevant information, compromising on factuality.

1. Evaluation Criteria for AI

• Consistency: Similar inputs should produce similar outputs.
• Factuality: Responses must be based on correct, verifiable data.
• Many models, including ChatGPT 3.5, have been found to generate 55% fabricated references, as per 2023 studies.

2. Real-Time Updating Challenge

• Experts argue no model can avoid hallucinations entirely unless updated in real-time with global data—an unattainable ideal at present.
• Claude AI creator Dhiraj Chattejee says: "You can’t fix hallucination without fixing how the AI is trained."

3. Benchmarks & Model Testing

• AI is tested using standardized benchmarks. But some models like ChatGPT-4 may have been “contaminated” by benchmark datasets.

• Chinese researchers found that benchmark-based learning can give misleading performance impressions.

• AI’s growth raises trust, transparency, and governance challenges.

• Hallucinations limit the use of AI in critical fields like healthcare, legal systems, education, and policymaking.

• Bias in training data and lack of contextual learning still plague modern AI.

• AI oversight and regulation will be crucial as hallucinations can spread misinformation at scale.

• India and global institutions must adopt policies on AI transparency, auditability, and explainability.

• There is a need to move toward domain-specific AI, trained with reliable, targeted datasets.

Mains Mock Question:

“With the growing adoption of Artificial Intelligence, hallucinations pose serious challenges. Discuss the causes, implications, and regulatory measures needed to ensure AI accountability and factual reliability.”