My college sophomore recently returned to campus, moved into her dorm, and set up her room. This triggered memories of my own experience decades ago, during which I remembered a little tradition I had: Each late August, organizing and decorating my new room, I’d listen to the same artist: Elvis Costello.

You may not know (or like) Britain’s New Wave poster boy of ‘nerd chic’, but Elvis Costello was a staple of my teens and twenties for his varied compositions and lyrical wordplay. And my favorite of his albums was and still is the incessantly punning Trust. I can vividly recall lugging a stereo upstairs into my dorm, plugging in my CD player, and unpacking while I blasted that album’s 14 tracks from beginning to end.

This fall, Trust is very much on my mind again–not the album, but rather the idea. I’ve spent recent months trying to forecast AI’s impact upon my company, CredSpark, upon our clients, and the knowledge economy verticals in which we operate. In the oceans of content about AI–running the gamut from breathless hype to near-terror–I’ve read & heard too little nuanced, deliberative discussion on the following questions: When and how much can we trust AI, and how will our level of trust influence AI’s growth & impact?

By my non-scientific estimates, 80% of AI coverage is about how fast & easy it makes knowledge work. But ‘fast’ and ‘easy’ are not synonyms for ‘trustworthy’. In fact, the speed and ease of something often causes us to glide right past questions of trustworthiness. Big Tech obviously benefits if we simply assume their AI is trustworthy, and let their various assistants draft our emails, transcribe our online meetings, and (they hope) plan and book our vacations. On the other hand, myriad stakeholders want us to distrust AI, out of concerns for the loss of human jobs, threats to national security from non-American AI models, or simply to get fear-based clicks on their content.

Neither the ‘AI Optimist’ nor ‘AI Pessimist’ point of view is particularly helpful. So the time has come for us to begin thinking in a balanced way about AI’s trustworthiness. Whether we’re using AI as consumers, or in our jobs, or as organizations looking for ways to leverage AI to create more value for our clients and stakeholders, we need to think a lot about when and how to trust AI.

For most of today’s consumer uses of AI, the stakes are relatively low. And when the stakes are low, it’s easier to trust. A marketer using AI to generate email subject line ideas, doesn’t incur much risk. But as AI becomes integrated into products and services, as we’re doing at CredSpark, I feel it’s critical that we tackle head-on questions of AI’s trustworthiness, both today and into the future.

AI and The Trust Factor: How Do We Think About It?

I’d offer that there are 3 ways to think about whether and how much we trust AI:

• Trusting AI’s Accuracy
  Trusting AI’s Intent
  Trusting AI’s Benefits

Each part of this series will allow me to unpack these ideas in order to (hopefully) provoke your own thinking. In this Part One, I’ll focus on the question of Accuracy.

Trust In AI’s Accuracy

There’s a good chance you remember the first time you used ChatGPT. Take a moment to cast your mind back there. What did it feel like?

As I sat with my laptop at our dining room table, the automated wordsmithing was so jaw-dropping, I didn’t think twice about how trustworthy or accurate it was. Then I started noticing its tendency to make up facts.

At first, the hallucinations were fairly easy to spot if you had some level of prior knowledge about the topic ChatGPT was writing about. Like me, you may have quietly chuckled at them, and felt reassured that only children or those not reading closely were likely to be fooled. But then ChatGPT and competing gen AI models began relentlessly improving, to the point where it became much harder to tell if they were producing knowledge or nonsense.

When you suspect a human is spouting BS, you can ask them, ‘Are you being serious?’ If you want to challenge them further, you can ask some version of, ‘How do you know what you’ve just said is accurate?’ Asking a person to explain their thought processes is asking them to engage in what educators call metacognition, i.e. thinking about your own thinking processes. I’d argue that metacognitive skill correlates with accuracy–if you’re thinking about your thinking, you’re probably more likely to catch your own mistakes, right?

The late 2024 introduction of ‘reasoning model’ LLM’s like GPT o3 was the dawn of something resembling metacognition in generative AI. Suddenly, users could follow the ‘thinking’ process of the models, reported step by step, as they took time to process our questions, rather than blurting out answers as quickly as possible.

What Will Increase Trust In AI’s Accuracy?

Assessing how much we can hope for AI’s accuracy to improve starts with grasping the basics of how they work. It’s a fascinating topic, and I’ve only a layperson’s understanding. But the most important thing to know is that LLMs are basically giant mathematical models of statistical probability–when prompted with a question into a chatbot, they predict which words should appear together with other words in the response. They can do this because they’ve analyzed billions of words of text written by humans, and spotted patterns in how words combine. This pattern recognition is a key part of training a model.

For nearly 3 years, since the launch of LLMs, the labs creating them have been focused on models trained on ever-larger sets of data, using ever-larger pools of computational power (i.e. server farms). Based upon recent critiques of GPT 5’s accuracy, some experts claim that we may be reaching the natural limits of how accurate LLMs can get under their current design.

But what should give us optimism that AI is on a path to generating more accurate results is the growing adoption of Retrieval Augmented Generation (RAG) and more generally, the connection of LLMs to proprietary databases. For those not familiar with the idea of RAG, it’s basically to combine the value of specific, curated, non-public sets of data (customer lists, product specifications, etc.) that businesses currently use with the conversational interface of an LLM. This allows you to ask questions in natural language which can best be answered with sources better than ‘what Redditors have written.’

The moment I first heard about RAG, I began to think ‘OK, now AI will finally become useful for business purposes.’ No longer will product managers need custom database queries written by software engineers; instead, businesspeople can directly ask questions of their own data, with a higher degree of confidence in the accuracy of that data. More accuracy in the results generated by AI means more trustworthiness, both for those inside of businesses and, when RAG-enabled search turns into products, for customers.

What Can You Do To Improve AI’s Accuracy?

In short: Embrace the RAG mindset, as we’ve done at CredSpark. ‘RAG’ itself is a quickly-evolving set of practices, the specifics of which are likely to continue to change. Rather than focusing on one specific approach, I suggest reframing your approach to AI: If you’ve not already done so, plan to evolve from off-the-shelf generative AI providers (even the paid offerings) and start connecting your LLMs to whatever data you’ve got that could better-inform the LLM’s output. For example, if your LLM is linked to your customer database or your product’s knowledge base, you can ask questions of a chatbot which will draw heavily from the content in those data sources.

You’ll obviously want to be mindful of your data protection commitments to your customers. But once you’ve invested the time to design a safe, compliant solution, the accuracy–and therefore business value–of your AI-powered services will increase rapidly.

Elvis Costello was ahead of his time when he sang, on Trust, “Pretty Words don’t mean much anymore.” Don’t be the gen AI user who settles for pretty words out of a chatbot; make sure you’re generating data-linked, accurate, privacy-protected results. And ensure whatever you’re building benefits you, not just big tech.

In the forthcoming Part Two of this series, we’ll explore the next important topic about trusting AI, namely, how to trust the Intent of those designing and using AI tools.