Building Truth in a World of Generative AI

Truth Matters – How Much Can you rely on AI

In 1960, Norbert Weiner published an article called ‘Some Moral and Technical Consequences of Automation’. In it, he discussed, that as machines learn, they may develop unforeseen strategies at rates that baffle their programmers. He highlights huge benefits and potential pitfalls of a world with self-learning ‘machines’. Such strategies, he thought, might involve actions that those programmers did not “really desire” and were instead “merely colourful imitation[s] of it.

As artificial intelligence (AI) continues to evolve at an accelerated pace, it has permeated nearly every aspect of our lives with its incredible potential to analyze data and uncover patterns that are invisible to the human eye. However, as the technology becomes increasingly sophisticated, it raises important questions about trust and safety. Among these concerns is the issue of truth verification in AI, especially when it comes to highly advanced language models like BARD and GPT-4.

In this blog, we will talk about some experiments we ran with these algorithms and the implications of their truth verification challenges, and the potential dangers posed to people who rely solely on AI-generated information.

BARD and GPT-4 (the next iteration of OpenAI’s widely acclaimed GPT-3) are state-of-the-art AI language models that have set a new benchmark in natural language processing (NLP). They can generate highly coherent text and are capable of understanding context, which makes them popular for a variety of applications, including virtual assistants, content creation, programming, and translation, to name just a few.

While BARD and GPT-4 have undoubtedly revolutionised NLP and shown immense promise in numerous domains, they also present a significant challenge: truth verification. Since these models are trained on massive data sets, they inevitably absorb biases, misconceptions, and falsities present in the data. Can we rely on them in their entirety?

We asked Google to reveal the trending topics for the day:

Let’s double down on Rishi Sunak. We decided to experiment with BARD and ask it why Rishi Sunak was trending. Here’s what it came up with

Well of course that can’t be right.

We know BARD is still improving and ChatGPT by itself can’t access intelligence till after October of 2021.

We then took a look at what Microsoft Bing had to say, given it is integrated with GPT-4 (one of the most advanced models till date) and thought of asking it what it had to say

Far from the present day situation and context of why he is trending.

A couple of Google Searches and reading articles later, we found reasons of him trending

Reasons for trending: Stephen Flynn challenging Rishi Sunak on the coronation, doctors voting on NHS strike and Zelenskyy’s UK visit.

Why are the most sophisticated AI models failing?

Large models of today can generate outputs that may appear well-structured and coherent but might not necessarily be accurate or true.

Furthermore, these models lack a built-in mechanism to validate the information they generate, making it even more difficult for users to discern the trustworthiness of AI-generated content.

The lack of truth verification in these sophisticated AI language models is far from trivial. With the rapid spread of misinformation in today’s digital age, AI-generated content that contains misleading or false information can exacerbate the problem and have detrimental consequences.

Take a look at our recent use case. With the intelligence of all the world’s information, on a topic as large as the UK Prime minister, results aren’t accurate or real and could be misleading to a massive extent.

Experimenting with

At we have long been pioneering experiments with LLMs and AI. We believe in building systems that promote transparency and empower marketers and decision makers with data driven insights and intelligence with verified insights.

Taking a look at what identified as the reason for Rishi Sunak trending:

Sounds pretty accurate. What we are doing differently here is that we run our intelligence across multiple models, feeding it with verified and accurate information and training it to provide the most optimal output.

Experimenting with Advanced Reasoning

We are told AI models have high reasoning skills and advanced analytical capabilities.

We thought of putting this to test.

We asked these AI behemoths if we can leverage this topic for a brand called Firstcry (an ecommerce website dedicated to all things for babies and infants).


Discounts to his supporters? Sounds fun and “doable”.

Let’s probe Bing

Pretty vague and non helpful if you ask me.

Let’s see what has to say

Our engine tells you it is a bad idea to leverage this topic and provides a BAD score of 2/10.

Short. To the point. Most importantly – TRUTHFUL

Key Learnings & Takeouts:

  1. Individuals who depend on AI-generated information may be misinformed or misled, leading them to make misinformed decisions or form incorrect beliefs.

  2. We can’t solely rely on “reasoning” of AI engines.

  3. Ironically, AI systems are black boxed and we can’t understand their ‘logical conclusions’.

  4. This learning is not just limited to AI agents and Large Language Models but holds true for “black-boxed” advertising platforms as well.

In conclusion, Truth verification is a challenge in the ecosystem. Consumers of intelligence need to be cautious of the output they get from black boxed platforms. Always choose a platform that is truthful and enables you to make important decisions better and faster. Choose solutions that you can rely on and you are trustworthy.

Abhinay Bhasin, Head of Product Marketing, ProfitWheel Inc.

Last updated