When a Swiss insurer offered me double the coverage in response to a letter written by ChatGPT, OpenAI's chatbot, I was taken aback. My work focuses on technology investments that typically have a “J” shaped return curve. With ChatGPT, it was straight up, an “I” of immediate return.
I've never seen a technology with so much potential be so accessible and so easy to use. Products like Microsoft 365 Copilot and Google Workspace with ChatGPT rival Bard, will make generative AI even more pervasive. Considering the approximately US$ 1 trillion available to venture capital funds and the five American technology giants, there will be plenty of investments for launching new products.
Ethical and legal issues are crucial and deserve careful debate, but the opportunities that generative AI offers have only begun to be explored. The success of ChatGPT, which reached 100 million users in just two months, according to an analysis done by UBS, is an example of this. But how to use this technology in a corporate environment?
OpenAI positions ChatGPT as a consumer service. Information shared with the chatbot may be stored and used by the company. This is not an environment to work with sensitive data.
However, it is still possible to use it in situations that do not involve confidential information, such as writing a column for a newspaper. OpenAI also offers another interface: the API platform with its Playground.
APIs allow communication between systems, and the ones from OpenAI offer privacy terms that are more appropriate for professional use. Currently, any submitted data is deleted after 30 days and is only used to monitor misuse or abuse of the technology, but it is not used to train its AI models. The Playground, in turn, provides a user interface to their APIs directly from a browser.
In Playground chat mode, the privacy terms are the same as with the APIs. In addition, you can adjust various parameters of the AI models available. Want more varied answers to the same question or expand topics in the answer? Just move the controls on the screen. Another difference to ChatGPT is in the billing, it is based on usage instead of a flat fee.
With the release of plugins, OpenAI has created its own app store, expanding access to its AI to the world. This service has limited access to the public, and its impact is yet to come. OpenAI and University of Pennsylvania researchers estimate that 49% of American workers will have at least half of their tasks affected by the integration of these new technologies. In a study published on March 27, they warn that, in general, the higher the salary, the greater the person's exposure.
In my case, I provided a list of ideas and asked ChatGPT for a formal text with market references justifying an increase in the premium to be paid by the insurer. The letter appeared in an instant, assertive and confident. And that's where the danger is, in the overconfidence.
As impressive as these natural language systems are, they don't understand what they are saying, or not in the same way we do. These programs are trained to discover the frequency of associations between words, expressions, and images, providing the most likely answer to the sequence received in a question or prompt. Based on existing data, they carry the biases of our reality and risk providing answers that belong to third parties’ intellectual property. In some situations, they invent answers or “hallucinate”, in AI jargon. On the blog AI Weirdness, author Janelle Shane presents amusing cases where the AI gets it wrong or creates non-existent “facts”.
Before using any response given by a generative AI system like ChatGPT, we must verify the veracity of the information. This raises the question of whether these systems actually increase productivity. If you have to check everything they produce, maybe it's easier to work without them and do it right the first time.
When I asked for the letter, I still wasn't aware of the tool's hallucination risks, but through a mixture of habit and luck, I validated one of the studies and removed the others for brevity. I'm not one to get it right the first time, and I always proofread my texts several times, so this did not have an impact on my experience. On the other hand, having obtained a fairer value from the insurer left a delicious taste of victory in using ChatGPT.
The fact that AI sometimes hallucinates doesn't negate the value it creates when it gets it right. The difference is often in how you ask your questions. It is important to create a context that is clear enough for the system to respond correctly, a practice called prompt engineering, one of the new jobs created by AI.
In our work routine, which is focused on "due diligence" of mergers and acquisitions,
We started using technology to accelerate some of our writing, as I did with the insurance company, but we quickly moved to integrate the APIs into our tech due diligence systems. An example is our Osint (Open Source Intelligence) process, in which we create an initial overview of companies that are potential acquisition targets for our clients by using public information sources, such as job advertisements and articles. With the use of AI, we were able to automate the report generation and focus on validating or correcting the data, cutting its time and effort by more than 50%.
Generative AI is trippy in many ways, it can blow our minds with its capabilities but it can also hallucinate. The Artificial Intelligence Index Report 2023, from Stanford University, argues that we need to understand it with more nuance and depth, as it raises complex ethical and legal issues but also helps us to solve significant challenges such as the creation of new antibodies, developing hydrogen fusion, or optimizing energy use, not to mention the productivity gains it can bring. With the accelerated growth of these tools' use and abilities, we must experience them to understand their impact in practice. In the age of artificial intelligence, our creativity has become even more important.