Perplexity AI's tips on getting the best results from itself.

AI's self-assessment of its limitations.

What sort of errors can you make and what should the user do to minimise these. For example, I've noticed that you are not always great at counting and that you sometimes need to be reminded of key principles, particularly related to the implications of breaching section 1082 of the Companies Act 2006.

Answer

As an AI, I strive to provide accurate and high-quality information, but there are certain limitations and common error types you should be aware of. Here’s a summary of the kinds of errors I can make, and practical tips for users to minimize them:

Common AI Error Types

1. Factual Inaccuracies

• Description: Sometimes, I may provide information that is outdated, imprecise, or simply incorrect, especially if the topic is highly specialized or rapidly evolving.

• Example: Misquoting a law, misreporting a statistic, or summarizing a study incorrectly.

2. Counting and List Errors

• Description: I may miscount items in a list, skip points, or repeat information, especially in longer or more complex responses.

• Example: Listing “five” benefits but only providing four, or duplicating a bullet point.

3. Contextual Misunderstandings

• Description: If the prompt lacks context or is ambiguous, I may misinterpret your intent or provide an answer that doesn’t fully address your needs.

• Example: Giving a general overview when you wanted a specific legal analysis.

4. Overgeneralization or Oversimplification

• Description: I might gloss over nuances, especially in complex legal, medical, or technical topics.

• Example: Summarizing a law without highlighting important exceptions or implications.

5. Missing or Incomplete Citations

• Description: Sometimes, I may not provide sources, or the sources may not be directly accessible or verifiable.

• Example: Referring to “a recent study” without a link or citation.

6. Legal and Regulatory Nuances

• Description: Laws and regulations (like section 1082 of the Companies Act 2006) can be complex, and I may miss subtle implications or recent amendments.

• Example: Failing to mention the latest case law or regulatory guidance that affects interpretation.

How Users Can Minimize AI Errors

1. Be Explicit and Precise

• Tip: Clearly state exactly what you need, including the level of detail, jurisdiction, date range, or specific principles to consider.

• Example: “Summarize the implications of breaching section 1082 of the Companies Act 2006, including recent case law.”

2. Request Verification or Sources

• Tip: Ask for sources, citations, or links to authoritative references, especially for legal, scientific, or technical information.

• Example: “Can you provide the official text or a reputable commentary on this section?”

3. Double-Check Critical Details

• Tip: For important facts, numbers, or legal points, verify the information with trusted primary sources or ask for a summary and then check the original documents.

4. Prompt for Corrections or Reminders

• Tip: If you notice an error (like a counting mistake or missing principle), point it out directly and ask for a correction or clarification.

• Example: “You listed only four points but mentioned five. Please review and correct.”

5. Iterative Questioning

• Tip: Use follow-up questions to clarify, deepen, or correct the initial response.

• Example: “You didn’t mention the penalties for breaching section 1082. Can you elaborate?”

6. Highlight Key Principles or Requirements

• Tip: Remind me of key principles or legal requirements you want included in the answer.

• Example: “Please ensure you address the fiduciary duties related to this section.”

7. Use Structured Requests

• Tip: Ask for information in a structured format (e.g., tables, bullet points, step-by-step analysis) to reduce ambiguity and make it easier to spot omissions or errors.

Summary

Error Type: How to Minimize

Factual inaccuracies: Ask for sources; verify with trusted references

Counting/list errors: Request structured lists; ask for review if numbers don’t add up

Contextual misunderstandings: Provide explicit context and clarify your needs

Overgeneralization: Ask for nuance, exceptions, or detailed analysis

Missing citations: Request citations or links to authoritative sources

Legal/regulatory nuances: Specify jurisdiction, date range, and ask for latest updates

In short:
Treat my responses as a helpful starting point, but always double-check critical information, especially for legal, medical, or high-stakes topics. Your feedback and precise instructions are key to getting the best results!

If you have a specific scenario or want to test how I handle a particular legal principle, feel free to ask!