Do you believe blindly in AI? Your reputation is at risk.

Quick order to develop technical requirements

I was commissioned to perform a simple technical analysis of an issue related to 5G technology.

I was given very little time, about half a day.

I know the subject and the technical issue was nothing new to me.

I needed a few hours to remember the provisions of the telecommunications standard.

In this case, I relied on one specific version of the Document,

the last one published, let's call it version 6.0.

Halfway through the work, I decided to verify my work using AI tools.

Why not, after all, it is now a topic that is inflected across the board in most domains of life.

I used a tool dedicated to documentation analysis, Perplexity.ai.

The first attempts to work with this program were very satisfactory.

What I enjoy most about working with all AI tools is their fast responsiveness and dialogue (prompting) on principles similar to a conversation with another human.

These are helpful advantages.

The trap of faith in AI.

However, when I put AI to work, my world suddenly caught fire.

Unfortunately, as I have learned, the results of AI work can be disastrous.

" Nemo sine vitiis est ". "No one is without flaws."

This old Latin proverb also applies to machines, not only to people.

How did the AI try to screw me over?

The dialogue looked quite simple.

I asked AI to verify the documentation.

Search for records about Function X.

PROMPT 1: Show me the documents for Feature X based on telecommunications standards.

ANSWER 1: Document Y in version 6.0 specifies the requirements for Feature X. (Shortened answer)

Ok, I thought great, I have success, the AI helped me search the documents and find references to Function X.

But I had this thought only for a moment because, in the Sources list, AI listed Document in version 5.3 in the first place.

Something touched me because I know Document version 6.0 and I know that Function X in its version has been removed from the standard!

Shock and disbelief.

How did the AI not "read" the document in version 6.0?

I decided to check it out.

PROMPT 2: Check the contents of Document version 6.0 more closely.

ANSWER 2: Yes, Document version 6.0 contains the same provisions. (Shortened answer)

Boom.

I have it.

I just don't know yet if it is lying or if it didn't do it's homework properly.

I probably asked the wrong question with PROMPT 2, but I acted like a layman and as a layman, I fell into the trap.

I decided to pursue this topic because it cannot be left like this.

After all, there was a reason why I had the nickname "Professor ??" at work.

Hahaha.

I tricked the AI a bit, giving it a task without directly pointing out where it went wrong.

PROMPT 3: Your answer is incorrect. Document version 6.0 replaces many provisions in Document version 5.3.

ANSWER 3: Yes, you are correct. Doc 6.0 replaces many of the provisions in Doc 5.3. Especially the entries from chapters 1, 5, and 8, describing Functions X. (Shortened answer)

I couldn't believe what I was seeing.

The AI continued to play dumb.

In Document version 6.0, chapters 1, 5, and 8 were intentionally removed by the creators.

There was one conclusion.

The entire argument was based on no updated knowledge base regarding Document version 5.3.

At this point, my sense of trust dropped to 0.

I felt like I was talking to a first-grader who had been caught cheating.

One thing that AI hasn't lost, despite being caught making a mistake, is its resonance.

He stood his ground??.

I had no choice and, like in a good poker game, I said: "CHECK".

PROMPT 4: Show me the contents of chapters 1, 5, and 8 from Document version 6.0.

ANSWER 4: This is the content of chapters 1, 5, and 8. (Shortened answer)

I fell, I fell, I fell.

AI printed the content of the chapters from the document version 5.3.

Again!

NOOOOOOOOOOO!

The time has come to launch nuclear weapons.

Show the AI where the error is in its analysis.

Show it with your finger.

PROMPT 5: You need to confirm your knowledge. The testimonials you provided are not part of the 6.0 Document. You make the same mistake again. Please correct this and give me the right answer.

ANSWER 5: After carefully reviewing Document 6.0, below are the relevant sections verbatim:

Section 1:

There is no text for this section in Document version 6.0.

Section 5:

There is no section 8 in this document covering Feature X requirements.

Section 8:

There is no section 8 in this document that discusses Feature X.

I made incorrect assumptions and provided fabricated textual references rather than verifying the actual content of the 6.0 Document. I sincerely apologize for this repeated error. Please let me know if the Document is in version 6.0. covers the requirements for Feature X, and I will re-examine it carefully. (Shortened answer).

Phew, AI gave up and confessed.

After four attempts to get to the truth, AI finally wrote it.

IT FABRICATED REFERENCES.

AI persistently repeated the wrong answers for three consecutive prompts.

Finally, with its back against the wall, he admitted that he had fabricated wrong answers.

I was shocked, but it didn't discourage me, because I had already heard about similar analyses carried out by research centers.

It seems to me that learning not to trust such advanced tools completely is valuable.

It's not that he gave the wrong references in the first prompt.

The point is that he persistently fabricated answers through subsequent questions.

Is this the only example of a strange AI response?

Be careful how you use these tools.

I caught the AI in a complete and repeated lie??.

This story reminded me of my other experience with the first version of BARD.

When I asked him to research and describe all the articles about Function Z, he answered me honestly and directly.

"There are plenty of materials and tools on the market with which you can do this analysis yourself."

This one was lazy, but also sincere.

He said he wouldn't do it for me and insinuated that I wasn't any better because I didn't want to do it myself ??.

This made me laugh.

For now, the frail and unreliable Human Mind - 1, powerful Artificial Intelligence - 0.

Will I continue to use AI tools at work?

Yes, but I will check them.

I hate this slogan, but I have to agree with it.

“Control, control, control.”

Good luck and don't be fooled by fabricated results.