What If AI Content Was Really Smart?

A client asked me a few years ago to teach a content AI how to find good material for a newsletter that was sent to over 20,000 C-suite leaders.

At that time, I was putting together 20 well-written business pieces from a lot of different sources. Instead, my client wanted the content AI to choose the articles. The end goal was for the newsletter to be totally automated.

The end result wasn’t great. The AI could find articles that were similar to ones that people had read before, but we couldn’t teach it to be smart, which means we couldn’t teach it to understand the vagueness of a new idea or a fresh way of talking about it.

My client cut the cord on the AI project and then the magazine itself in the end. But I’ve been thinking about that experience as large language models (LLMs) like OpenAI’s GPT-4o continue to get more attention from the public.

I wonder if we would have been able to find “good” stories more easily today if we had used an API into GPT-4o.

Content AI tools like ChatGPT and Jasper.ai are based on GPT-4. These tools are very good at understanding language prompts and writing clear text very quickly on almost any subject. But there is a bad side to content AI: the smart content they make can be boring, and they make stuff up all the time. No matter how fast and fluent they are, the big language models of today don’t think or understand like people do.

What if they did? What if AI makers found ways to get around the problems that content AI has now? In other words, what if content AI was really smart? They are already getting smarter in a few ways. Let’s talk about how content professionals can use these improvements in content AI to achieve their goals.

Five ways that content AI is getting smarter

It helps to go over how big language models work again in order to understand why content AI isn’t really smart and how it’s getting smarter. As you go through a phrase, GPT-4 and “transformer models” like Gemini by Google, Claude by Anthropic, or Llama by Meta look at all the data (words) and how they relate to each other. These are examples of deep learning neural networks.

The AI developers trained them with web content, which gave them a lot more training data with more factors than before. This made the outputs more fluent for a wider range of uses. That being said, Transformers don’t know what those words mean in real life. The models can only see how they are usually put together in words and how they connect syntactically.

So, generative AI today works by guessing what the next word will be in a string based on millions of previous phrases that are similar. This is one reason why users of big language models often have “hallucinations,” or information that they made up, or get wrong information. These tools are only making up sentences that look like sentences they have seen before in their training data. All kinds of mistakes, useless details, debunked facts, and false equivalencies will show up in the generated language if they are present in the teaching language. A lot of AI experts even think that dreams will happen.

Still, you can lessen them. As this clever “leader board” for hallucinations shows, today’s big language models dream less often than their predecessors did. Also, both data scientists and users can get rid of them in a number of ways.

Answer #1: AI content suggestions
Prompting is something that everyone who has used an AI app knows about. For the most part, you tell the tool what to write and sometimes how to write it. There are easy questions like, “List the good thing about having AI write blog posts.”

You can also make the prompts smarter. You can give the content AI an example paragraph or page of text written in the voice and style of your company and ask it to write subject lines, social media posts, or a new paragraph in the same voice and style.

Prompts are one of the best ways to set rules that limit what content AI can do. If you keep your prompts focused, clear, and specific, the AI is less likely to write copy that isn’t on brand or isn’t based on facts.

A type of rapid engineering called retrieval augmented generation (RAG) is also being tried out by businesses. Users tell the model to answer the question using a certain source of knowledge, which is usually not part of the original training set. This is called RAG-enhanced prompts.

RAG doesn’t stop dreams completely, but it can help content experts find mistakes because they know what content the AI used to answer.

If you want to learn more about prompting methods, read this article for content marketers about writing AI prompts or this article about researcher Lance Elliot’s nine rules for writing prompts that will stop hallucinations.

Solution No. 2: A “chain of thought”
Think about how you would solve a math problem or show someone the way in a city you don’t know where they are going that doesn’t have any street signs. You would probably divide the problem into several steps and use deduction to solve each one. This would lead you to the answer.

A similar method is used in chain of thought prompting to break down a thinking problem into several steps. The goal is to get the LLM to write text that looks like it was thought through in a reasonable or sensible way.

Scientists have used chain of thought methods to help LLM do better on math problems and more difficult tasks, like inference, which people do naturally because they understand language in context. Experiments have shown that users can get more accurate results from LLMs when they are given chain of thought hints.

There are even experts working on adding pre-written prompts and chain-of-thought prompts to LLMs so that most people don’t have to learn how to use them.

Solution #3: Making content AI work better
Fine-tuning means taking a large language model that has already been trained and teaching it to do a specific job in a certain field by giving it relevant data for that field and getting rid of irrelevant data.

A fine-tuned data language model should be able to recognize and generate language as well as the original, but it should focus on a more specific context to get better results.

There are a lot of cases of fine-tuning for things like writing legal documents, financial reports, tax forms, and more. A company can make a new tool that can write smart content with fewer dreams by fine-tuning a model with writings from court cases or tax returns and fixing mistakes in the results it generates.

It might not make sense for these government-controlled or regulated fields to use technology that hasn’t been tried much. For example, a Colombian judge is said to have used ChatGPT to write his decision brief (without fine-turning).

Answer #4: Creating specialized models
Many people think that fine-tuning a model that has already been trained is faster and cheaper than making new models. But that’s not the only way. If they have enough money, researchers and tech companies can also use transformer models to make language models that are specific to certain jobs or domains.

For instance, researchers at the University of Florida and Nvidia, a company that makes AI technology, worked together to create a special large language model for health that can be used to look at and evaluate language data in the electronic health records that hospitals and clinical practices use.

As a result, GatorTron was made, which is said to be the most well-known LLM intended to look at what’s in clinical records. The group has already made a similar model using fake data, which takes away the privacy concerns that come with using AI content that is based on real medical records.

In a recent test where the model was used to make doctor’s notes, AI-generated content was 50% of the time hard for humans to tell was produced by AI.

The main screen of the prompt library for Anthropic for a piece on Content AI. This picture has a lot of text and basically shows a search bar with choices.

Solution #5: Features that can be added
Making material is often part of a bigger process in the business. Some developers are adding functionality on top of the material to make it more useful.

Some experts are working on prompting add-ons so that regular users don’t have to learn how to prompt well.

That is just one case. Another one is from Jasper. The improvements they’ve made to Jasper for Business are a clear attempt to get high-level contracts. One of these is a user interface that lets people set the “brand voice” for their company and use it in all the copy they write. People can also use Jasper in business apps that need text by creating bots that work with Jasper.

Another company, ABtesting.ai, combines language generation with web A/B testing to try out different versions of web copy and calls to action (CTAs) and find the best one.

What to Do Next to Use Content AI
The methods I’ve talked about so far are extensions or improvements to the basic models. But as the field of AI continues to grow and change, researchers will make AI that is more like humans when it comes to thought and reasoning.

The Holy Grail of “artificial generation intelligence” (AGI)—a type of meta-AI that can do many different kinds of computing tasks—is still alive and well. Others are looking into ways to make AI capable of abstraction and comparison.

The lesson for people whose life’s work is to make good content: AI will keep getting smarter. We can also “get smarter,” though.

I don’t mean that people who make AIs try to beat them at jobs that need a lot of computing power. That being said, the AI needs help and direction for now. These are the main thoughts you should use when writing. And even if a content AI finds something new and original, it still needs people to value it and make it a top priority. That is, creativity and new ideas will always be in the hands of people. The longer we use those skills, the bigger our lead gets.

Source

Radiant TV, offering to elevate your entertainment game! Movies, TV series, exclusive interviews, music, and more—download now on various devices, including iPhones, Androids, smart TVs, Apple TV, Fire Stick, and more.