Though Generative AI or Large Language Models (LLMs) such as Google’s Bard or OpenAI’s ChatGPT have become commonplace in business and in copywriting, not is all as it seems.
If you spend enough time using these tools such as ChatBots or online content generation tools, the AI may produce results that are bizarre, make no sense, or are completely made up.
This phenomenon is known as “AI Hallucination” and can prove damaging for professionals in creative industries such as content writing and copywriting. Here’s what copywriters and other creatives can do about it.
How Generative AI works
Generative AI doesn’t “think” like we do – in a nutshell, the AI manipulates numbers to reach a desired outcome.
Though the language it generates is human readable and appears fluent, it is predicated on a series of tokens broken down into probabilistic models. If one provides a prompt “Write the next word in this sentence: The cat sat on the ___?”
The Generative Pre-Trained Transformer (GPT) will parse that phrase like a game of Family Feud – the survey – or information it is fed, will produce the response of the highest probability according to the data it can access. To do this, it lends certain words and phrases different probabilistic weights – The cat sat on the mat. “Mat” may be weighted with near 100% probability while “laptop” may be weighted with 10-15% probability.
It “learns” this through analysing millions, if not billions of words and sentences pulled from various online repositories.
The AI does not and cannot “know” the implied meaning of the words it generates in isolation, or even in sequence. It is manipulating numbers and algorithms to produce the “answer.”
What is an AI Hallucination?
Hallucination, a term from human psychology, refers to an individual seeing or hearing things that in reality, do not exist.
Likewise, AI hallucinations present incorrect, inaccurate, or logically inconsistent information as factually correct and complete information – and being confident that the returned result is indeed truthful.
The term “AI Hallucination” has been around since at least 2018, when Google authored a paper on Neural Machine Translation, where they said “NMT systems are susceptible to producing highly pathological translations that are completely untethered from the source material.”
Many Generative AI models are trained to provide answers at almost “any cost.” Given an AI is plugged into vast reservoirs of already-existing information at all times, it is hard to understand how an AI can “not know” a given answer when asked a straightforward question. Therefore, it simply produces something, rather than nothing.
In fact, OpenAI disclaims hallucinations by warning users that “ChatGPT may produce inaccurate information about people, places, or facts.”
Other terms for AI hallucination are “confabulation” which can appear in AI-generated video, upscaling, images, or audio as weird artifacts or odd interpretations of prompts.
Common types of AI hallucinations
The most common types of AI hallucinations that can creep into copy are factual errors or inconsistencies in the output.
The most famous example is from February of 2023. Google’s Generative AI Bard stated that NASA’s James Webb Space Telescope took the very first pictures of a planet from outside our solar system.
Astronomers quickly pointed out that this was incorrect.
It can also contradict itself and not be “aware” of the contradiction. You may ask the AI to write a 26-word poem with each word starting with a corresponding letter of the alphabet, and it returns 25 words or 27 words. Even when “challenged” it may insist that the output conforms to what the prompt has asked.
Other (more fun) hallucinations are when the AI goes completely off the rails and returns gibberish or nonsensical data. You may ask for an explanation of the water cycle, and it goes off on a tangent about 80s horror films. In French.
This can all happen due to the quality of data being collected, how it generates responses, how well a prompt is engineered, and plain old human error in the form of bugs.
Refer us and get $50!
〰️
Refer us and get $50! 〰️
AI hallucinations and copywriting or content writing
With the right prompts and skill, a LLM or Generative AI can produce massive amounts of content at scale.
However, the downside is, at least for the moment, is that AI generated content lacks the “Experience” ingredient in Google’s page ranking factor model, known as E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness.)
If a Generative AI “hallucinates” once or twice per 1,000 words, this can make even the most well-intentioned content run afoul of Google’s E-E-A-T parameters.
If a human website copywriter cannot determine whether the AI has hallucinated or not (e.g. through a lack of expertise on the topic being covered) it can wreak havoc not only for their SEO efforts but also for productivity.
Copywriters and editors will need to redline and fact-check each claim a Generative AI has made. For shorter copy and content (under 1000w), there is a reasonable chance that writing the post oneself would take less time than crafting prompts and checking facts.
How to prevent AI hallucinations
Though we can’t totally stop an AI from hallucinating, we can take steps to mitigate the risk of hallucinations occurring. This can come from writing better prompts or feeding it higher quality information.
Pre-train the AI
The “PT” in GPT means “pre-trained.” We may assume that a GPT is adept in all things but feeding it relevant and contemporary data first can help limit factual errors. These may come into play by asking a GPT to read through several hand-picked sources or databases before generating text.
Give it a role to play
Image generative AI such as DALL-E and Midjourney are called upon to “Imagine” images based on prompts; and GPTs can be given roles to play which can narrow its focus to certain types of data. For example, if you want an expert to determine whether the James Webb Telescope did indeed see an extra-solar planet, you can ask it:
You are an astronomer who works at NASA and concentrates on reading telemetry from the James Webb Telescope. Has the telescope ever seen a planet outside our solar system?
Exclude certain answers
Just like the dark days of the pre-Google internet, search operators were crucial to find relevant information. If you wanted to look up “bicycles” you may have had to Ask Jeeves (literally) “Bicycles NOT motorcycles”, which would exclude motorcycles from your results.
This could take the form of:
Provide a list of islands in the South Pacific which exist in the real world. Exclude all fictional mentions and return what can be seen on a contemporary (2010s onwards) map.
Forgive its trespasses
If the GPT has racked its neural networks and databases and come up with nothing, let it know that that is also an acceptable answer.
Does nothing unreal exist? If you do not know, say, “I don’t know.”
Copywriters: trust, but verify
As of 2023, hallucination is part and parcel of using Generative AI. As a copywriter or content writer who may look to GPTs for ideas, long-form content generation, or productivity improvements, we need to rely on the old political maxim: trust but verify.
There’s a high probability that the GPT output is factually correct, but we also need to cross-check with third-party reputable sources before publishing; lest you contribute to spreading misinformation unwittingly.
To get the best human copywriting and copyediting you can find online, contact me! No hallucinations, guaranteed.