AI “hallucinations” – those convincing-sounding but false answers – draw quite a lot of media attention, as with the recent Latest York Times article, AI Is Getting More Powerful, But Its Hallucinations Are Getting Worse. Hallucinations are an actual hazard once you’re coping with a consumer chatbot. Within the context of business applications of AI, it’s a fair more serious concern. Fortunately, as a business technology leader I even have more control over it as well. I can be certain the agent has the precise data to supply a meaningful answer.
Because that’s the actual problem. In business, there isn’t a excuse for AI hallucinations. Stop blaming AI. Blame yourself for not using AI properly.
When generative AI tools hallucinate, they’re doing what they’re designed to do – provide the very best answer they’ll based on the information they’ve available. Once they make stuff up, producing a solution that will not be based in point of fact, . Yes, recent models like OpenAI’s o3 and o4-mini are hallucinating more, acting much more “creative” after they don’t have a great answer to the query that’s been posed to them. Yes, more powerful tools can hallucinate more – but they can even produce more powerful and worthwhile results if we set them up for achievement.
For those who don’t want your AI to hallucinate, don’t starve it for data. Feed the AI the very best, most relevant data for the issue you wish it to resolve, and it won’t be tempted to go astray.
Even then, when working with any AI tool, I like to recommend keeping your critical pondering skills intact. The outcomes AI agents deliver will be productive and pleasant, but the purpose will not be to unplug your brain and let the software do all of the pondering for you. Keep asking questions. When an AI agent gives you a solution, query that answer to make certain it is sensible and is backed by data. If that’s the case, that needs to be an encouraging sign that it’s value your time to ask follow up questions.
The more you query, the higher insights you’re going to get.
Why hallucinations occur
It’s not some mystery. The AI will not be attempting to deceive you. Every large language model (LLM) AI is actually predicting the following word or number based on probability.
At a high level, what’s happening here is that LLMs string together sentences and paragraphs one word at a time, predicting the following word that ought to occur within the sentence based on billions of other examples in its training data. The ancestors of LLMs (except for Clippy) were autocomplete prompts for text messages and computer code, automated human language translation tools, and other probabilistic linguistic systems. With increased brute force compute power, plus training on internet-scale volumes of knowledge, these systems got “smart” enough that they might carry on a full conversation over chat, because the world learned with the introduction of ChatGPT.
AI naysayers prefer to indicate that this will not be the identical as real “intelligence,” only software that may distill and regurgitate the human intelligence that has been fed into it. Ask it to summarize data in a written report, and it imitates the way in which other writers have summarized similar data.
That strikes me as an instructional argument so long as the information is correct and the evaluation is beneficial.
What happens if the AI doesn’t have the information? It fills within the blanks. Sometimes it’s funny. Sometimes it’s a complete mess.
When constructing AI agents, that is 10x the danger. Agents are purported to provide actionable insights, but they make more decisions along the way in which. They executed multi-step tasks, where the results of step 1 informs steps 2, 3, 4, 5, … 10 … 20. If the outcomes of step 1 are incorrect, the error shall be amplified, making the output at step 20 that much worse. Especially, as agents could make decisions and skip steps.
Done right, agents accomplish more for the business that deploys them. Yet as AI product managers, now we have to acknowledge the greater risk that goes together with the greater reward.
Which is what our team did. We saw the danger, and tackled it. We didn’t just construct a flowery robot; we made sure it runs on the precise data. Here’s what I believe we did right:
- Construct the agent to ask the precise questions and confirm it has the precise data. Ensure that the initial data input technique of the agent is definitely more deterministic, less “creative”. You wish the agent to say when it doesn’t have the precise data and never proceed to the following step, relatively than making up the information.
- Structure a playbook to your agent – be certain it doesn’t invent a brand new plan each time but has a semi-structured approach. Structure and context are extremely necessary at the information gathering and evaluation stage. You possibly can let the agent loosen up and act more “creative” when it has the facts and is prepared to write down the summary, but first get the facts right.
- Construct a prime quality tool to extract the information. This needs to be greater than just an API call. Take the time to write down the code (people still do this) that makes the precise quantity and variety of knowledge that shall be gathered, constructing quality checks into the method.
- Make the agent show its work. The agent should cite its sources and link to where the user can confirm the information, from the unique source, and explore it further. No slight of hand allowed!
- Guardrails: Think through what could go improper, and construct in protections against the errors you absolutely cannot allow. In our case, that implies that when the agent tasked with analyzing a market doesn’t have the information – by which I mean our Similarweb data, not some random data source pulled from the net – ensuring it doesn’t make something up is a vital guardrail. Higher for the agent to not have the opportunity to reply than to deliver a false or misleading answer.
We’ve incorporated these principles into our recent release of our three recent agents, with more to follow. For instance, our AI Meeting Prep Agent for salespeople doesn’t just ask for the name of the goal company but details on the goal of the meeting and who it’s with, priming it to offer a greater answer. It doesn’t should guess since it uses a wealth of company data, digital data, and executive profiles to tell its recommendations.
Are our agents perfect? No. No one is creating perfect AI yet, not even the largest firms on this planet. But facing the issue is a hell of lots higher than ignoring it.
Want fewer hallucinations? Give your AI a pleasant chunk of top of the range data.
If it hallucinates, perhaps it’s not the AI that needs fixing. Possibly it’s your approach to making the most of these powerful recent capabilities without putting within the effort and time to get them right.
