The AI field is growing, and whether good or bad, people are doing more than talking about it; they’re using it more than ever. However, despite this increased use, I’ve noticed that, for some, their perception tends to alternate between false and too-high expectations of AI.

One case, in particular, was in 2021, Gartner placed natural language processing (NLP) at the top of its list of loaded expectations in terms of the Gartner hype cycle. As a result, many expected a potential “winter of AI,” so to speak. Yet, in 2022, we discovered the potential that we haven’t even touched on the true value AI could deliver.

Will there be a “winter of AI,” and are expectations bloated?

No, I don’t think so. As the past year has shown us, AI still has more to offer, a pocket of value that we have yet to see. I believe that while many people now accept that AI will be a transformative force—thanks to the fast democratization of large language models—our society hasn’t yet fully considered the actual changes it will make by lowering the barrier to access intelligence globally.

Progress in image generation, analysis, and computer vision—think autonomous driving—has leaped and bounded in the past year, and so has the progress in NLP, particularly in the natural language understanding (NLU) and natural language generation (NLG) aspects. We’re at a tipping point that will likely transform our world in the same way that the internet has.

Tipping point for AI

Today, we’re seeing the development of natural language processing through large language models, such as with the emergence of ChatGPT based on OpenAI’s large language model version GPT-3.

Astounding fact: ChatGPT’s growth in user adoption skyrocketed past one million users within a week of launching. In comparison, no other tech company has reached this feat in this short of a time frame. But the adoption rate is only part of it.

This advance has profoundly affected creative jobs because this might be the first time an AI generative system can create high-quality content. In public mode, users have tapped ChatGPT to do everything, from generating basic reports and ideas to writing lectures and producing code.

With a high adoption rate comes great opportunity. Any startup seeing this level of success could become the most funded project ever. And more, there’s revenue. OpenAI, as the example, could make one billion dollars by 2024, according to a report via Reuters.

On the other side of the same coin, however, there are greater risks due to AI generative system advancement. For example, with AI assistance, human hackers can develop more sophisticated phishing campaigns—hacking mechanisms based on social engineering.

Competition, specificity, and focus for AI advancement

Despite the risks, we still haven’t seen what’s yet to come with generative AI. GPT-4, for instance, is rumored to launch in 2023. I believe it will be a massive improvement over GPT-3, which is already mind-blowing.

And on the point of NLG and these large language models, there’s a lot that’s feasible in process automation. For context, creative content gets the most attention; it’s the area that makes more headlines. But I would also watch advancements in technical content and automated code generation, for example.

Process automation

Because of today’s AI advancements, it’s now possible for tools like ChatGPT to generate near-ready-to-use source code. That means instead of only being fun to play around with, these are becoming enterprise tools, making it possible for developers to automate technical tasks at scale.

NLP—specifically natural language understanding, which SESAMm works on—is not untouched by these applications. Many of these large language models can perform zero-short learning, which means NLU can be performed without pre-training, a huge advance in this industry. However, zero-short learning is insufficient for many advanced sentiment and ESG analysis tasks. We still need additional data sets to fine-tune the data for a specific purpose.

What does this mean for the natural language generation sector? Many startups—especially anything around chatbots—have folded, some just in Q4 of 2022. ChatGPT’s success means it’s solved and replaced the need for many of them, and basically, anything content creation on the B2C side has and will struggle.

Defensive edge

Otherwise, things are looking good in our sector. For example, at SESAMm, we’re focused on what I call “last-mile AI.” In our specific business application, you can’t bypass the need for a data set because we’re trying to attain a precise result for specific, often risk-related applications. Open-source large language models like GPT-3 and BERT can get you mostly there, and that’s fine for general purposes. But for “last-mile AI” applications, there’s a lot you can’t do without additional work.

And here lies what I think is one of SESAMm’s defensive edges: the “last-mile AI.”

Instead of finding ways to protect its algorithms, the AI business community would do better to defend its use cases because the algorithm’s value will decrease progressively. In contrast, the value of a use case’s purpose and the data set used to achieve the use case will grow.

Competitive edge

Computing power and the resources it takes to train large language models remain challenging to applications like OpenAI. It takes electricity, heat, and money to train these models, and AI has an environmental impact. So far, we’ve justified this cost in the name of optimization—meaning that we put in this extra cost upfront so that the likely efficiency will offset or reduce that cost later—but it’s still a cost to incur.

AI companies, especially those in the NLG space, will do well to find their competitive edges, areas optimized for a specific purpose like “last-mile AI.” Companies like OpenAI will likely continue to optimize their models for quicker responses but don’t necessarily have the problem of solving for a specific use case.

At SESAMm, for instance, a big challenge and expertise we developed in-house is inference time—or how quickly we can apply the model to an article or an individual sentence. Because we’re processing so much live content, the more time it takes to process—milliseconds multiplied by a billion—the more costly it is.

Our data lake currently holds over 20 billion articles, messages, etc., from over 14 years, and we add 10 million more daily. That’s a lot of content to analyze. But we make it so our clients can access the data within seconds.

The need to optimize models for fast inference and adapt to deep industry-specific use cases will remain one of the key reasons companies will have to continue re-training their own models. That doesn’t mean large language models don’t add value here. Their open-source versions simply become an impressive building block for any NLP application and accelerate the rate of innovation and productivity in the whole field.

My summary thoughts on AI for 2023

When Google launched BERT in November 2018, we quipped that Google had open-sourced this system as a joke because no one could put it into production because BERT was so big. Many companies didn’t have the computing capabilities to do anything with it at the time. Now we do.

This year, Google did it again; they released a model that’s even bigger than GPT-3. Of course, almost no one besides Google can put that model into production now. But my point is that there will always be computing, resources, and other challenges to making AI advancements. That’s why I think AI companies must focus on defensive and competitive edges.

Regardless of the challenges, I see good things happening in the NLU space being massively improved by large language models. I see improvements as we incorporate these models today compared to deep-learning models trained from scratch a few years ago. I also see a significant decrease in the amount of data we need to fine-tune results, reaching and focusing on the final client use case more quickly.

From a natural language generation perspective, I believe large language models will transform the world. And I’m really excited about this era because this transformation supports my deepest purpose, leveraging AI to accelerate innovative decision-making. We do this by giving decision-makers access to technology that analyzes research content, news, and discussions. And if we increase the rate of innovation or the quality of decision-making by 10% globally, the impact could be huge for all industries: healthcare, finance, fashion, you name it. Industry leaders can make better ESG and SDG choices that will affect our world on a grander scale.

2023 will be an exciting time for AI, specifically for NLG and NLU. Of course, we’ll continue to see AI innovations. But more importantly, leaders will have better insights to make better decisions, creators will create more—and more complex—content, and overall, the applications will become more specific to solving the needs of particular use cases.

Here’s to the new era of AI in 2023. Cheers!

About SESAMm

SESAMm is a leading NLP technology company serving global investment firms, corporations, and investors, such as private equity firms, hedge funds, and other asset management firms. SESAMm provides datasets and NLP capabilities through TextReveal^® to generate alternative data for use cases, such as ESG and SDG, sentiment, private equity due diligence, corporate studies, and more. With access to SESAMm’s massive data lake, comprised of 20 billion articles and messages and growing, its clients can make better investment decisions.

SESAMm recently hosted a webinar led by Lead Solutions Engineer Leo Shamash. The session focused on the critical role of Artificial Intelligence in identifying and managing ESG (Environmental, Social, Governance) risks and controversies, especially in private companies.

During the webinar, Leo Shamash shared insights on how SESAMm’s advanced AI technologies analyze millions of daily articles to provide accurate ESG risk assessments.

Why is this important for private equity firms? Because traditional methods of risk assessment are often labor-intensive and limited in scope. SESAMm’s AI-driven approach offers a scalable, efficient solution. The webinar also touched on SESAMm's extensive data lake comprising over 20 billion documents, making it one of the largest repositories for tracking ESG risks and controversies.

Watch the webinar replay now:

‍

Join us for our next webinar on November 15 at 4 PM Paris time/10 AM New York time, and watch Sylvain Forté share his insights into how artificial intelligence can help distinguish between genuine sustainability efforts and greenwashing. Book your spot.

Reach out to SESAMm

TextReveal's web data analysis of over five million public and private companies is essential for keeping tabs on ESG investment risks. To learn more about how you can analyze web data or request a demo, contact one of our representatives.

A familiar debate has followed ESG data for years. One camp argues that the field generates too much information for any human team to handle, so the work should be left to machines. The other argues that ESG judgments are too consequential to automate, so humans should review everything. Both positions contain a real concern. Neither describes how a credible rating is actually produced.

With the publication of its full Controversy Exposure Score methodology, now public and free to access following the entry into force of the EU ESG Rating Regulation on 2 July 2026, SESAMm is making the answer explicit. A trustworthy rating is not a choice between artificial intelligence and human expertise. It is the disciplined combination of the two, with each doing the part of the work it does best.

The Scale Problem Is Real

Teams that monitor ESG controversies rarely suffer from too little information. They suffer from too much. A single incident can generate dozens of articles within days, in multiple languages, across outlets of very different quality. Multiply that by a global investment universe and the volume becomes impossible to track by hand.

This is the part of the problem that machines are built for. SESAMm's pipeline ingests more than 10 million documents a day, drawn from an input layer of over 30 billion documents that includes licensed global news, public web and media feeds, NGO publications, and public regulatory and judicial filings. It screens controversies for millions of public and private companies, alongside infrastructure projects, state-owned entities and sovereigns. No analyst team could read at that scale, and none should try. Asking people to do machine work is how important signals get missed.

So artificial intelligence carries the load. Natural language processing and machine learning models, including large language models, attribute documents to the right entity, filter for genuine ESG relevance, and group related articles into discrete events and events into continuous cases. This is what allows a controversy that unfolds over weeks to be tracked as one developing story rather than a hundred disconnected headlines.

Why Scale Alone Is Not Trust

A system that reads everything will also, inevitably, misread some of it. SESAMm is direct about this in its methodology, because pretending otherwise would be the opposite of transparency.

Probabilistic language models can misinterpret a historical or hypothetical reference as an active controversy. Automated clustering can occasionally merge two distinct incidents or split one prolonged crisis into fragments. Model accuracy varies across languages, and lower-resource languages or heavily idiomatic content raise the risk of misclassification. These are structural properties of statistical systems, not bugs to be wished away.

This is precisely where scale stops being enough and human expertise becomes indispensable. A number that informs how capital is allocated cannot rest on automation alone.

Where Human Judgment Enters

SESAMm operates a dual-layer human quality-assurance process, and it runs every day.

At the first layer, a dedicated Research and Analytics quality-assurance team reviews data accuracy, both reactively, when a question is raised about a case, a score or a classification, and proactively, by reviewing generated alerts. Where an issue is confirmed, the correction, whether a reattributed entity, a corrected sub-risk tag or the removal of an irrelevant event, is applied at the source, logged, and the affected scores are recomputed on the standard daily cycle.

At the second layer, complex cases and recurring structural issues are escalated to the Methodology Lead, who can update the underlying training corpus so that a category of error becomes less likely in future. This is the detail that matters most. Human review is not a final rubber stamp on top of the machine. It is a feedback loop that teaches the system, so that today's corrections improve tomorrow's automated output.

The same expertise sits at the front of the process, not only the end. Analysts define the 44 ESG sub-risks, design how severity is assessed, and fine-tune the models. The methodology is a human construction that machines then apply consistently at scale.

A Division of Labor, Not a Contest

Seen this way, the old debate dissolves. The question was never whether AI or people should produce ESG ratings. The question is which part of the work belongs to which.

Machines provide reach, consistency and speed. They apply the same rules to every entity, every day, without fatigue or favor.

People provide judgment, correction and improvement. They decide what the system should look for, they catch what it gets wrong, and they raise the standard of the model over time.

The result is a rating that is both broad enough to cover the real world and rigorous enough to be relied upon. Scale without rigor is noise. Rigor without scale never reaches most of the companies an investor actually holds. The value is in the combination.

What This Means Going Forward

The EU ESG Rating Regulation asks providers to disclose how their ratings are built. SESAMm has chosen to disclose the full pipeline, including the role of AI, the points where it can fail, and the human controls that contain it. The aim is not to claim the technology is flawless. It is to show, in detail, why the output can be trusted anyway.

As artificial intelligence becomes more capable, the temptation to remove the human layer will grow. SESAMm's position is the opposite. The more powerful the models become, the more valuable the people who direct them, check them and teach them become. That is the architecture of a rating worth trusting, and it is now open for anyone to read.

To explore the full methodology behind the Controversy Exposure Score, visit sesamm.com/methodology.