Best Prompting Strategies For Overcoming The Exasperating Reverse-Curse Deductive Logic Weakness Of Generative AI

Are you smarter than AI?

Let’s do a fast test and see.

Suppose I tell you that the father of James Coburn is Samuel Coburn. I then ask you who is the son of Samuel Coburn. Can you answer that question? I would assume that all else being equal, you would quickly and without any hesitation reply that James Coburn is the son of Samuel Coburn.

Easy-peasy.

Turns out that the generative AI of today might not produce that correct answer. The AI might balk and insist that the answer is unknown. Or the AI might provide a made-up answer such as indicating that the son is perhaps Shazam Coburn or some other fictitious name. To clarify, there is also a chance that the AI will provide the right name. Like a box of chocolates, you never know what you might get as an answer.

The issue at hand is that the underlying computational mechanisms within generative AI are poorly suited to do the kind of logic involving deductive-style reasoning. Humans seem to be pretty good at deductive logic, by and large. When you are told that the father of James is Samuel, you seemingly create a relationship in your mind of an A-is-to-B nature, namely that Samuel is the father of James. You can then readily mentally figure out the reverse that consists of a B-is-to-A condition. Your mind comes up with the notion that James is the son of Samuel. Children can even do this (depending on their age and maturity).

We take for granted that deductive logic is a part of human thinking.

This is not so with most of today’s generative AI.

Generative AI is usually devised on a text-based pattern-matching basis that seeks to computationally determine what text is most likely to come next in a sentence or paragraph. If I feed in a sentence that the father of James is Samuel, the odds are that generative AI will pattern onto the aspects that the word “father” is followed by “James” and further followed by “Samuel”. Later on, a question about the father of James is somewhat easily answered by reusing that devised pattern and swiftly emitting that Samuel is the father.

Unless there happens to also be a sentence in the provided input that explicitly says that James is the son of Samuel, it is somewhat unlikely that the generative AI is going to have a suitable devised pattern in its internal structure of the deductive B-is-to-A concoction. No such pattern is automatically derived in most of today’s generative AI.

The generative AI will report that there isn’t any answer available to the question about the son of Samuel. Another possibility is that the internal linkages might point to some other name such as Shazam, perhaps based on other sentences of a similar look-and-feel, and mathematically spur the AI to indicate that the son’s name is Shazam.

People are often shocked and dismayed that generative AI might not be able to answer correctly such an easy question. The overall fluency of generative AI lulls us into thinking that GenAI is totally conversant and possibly on par with human interactivity. You must fight the strident urge to assume this. Generative AI is primarily a computational pattern-matcher and does not enter into the realm of sentience. Be very cautious when using generative AI and realize that the amazing appearance of fluency has lots of limitations and caveats that need to be soberly considered.

The aspect that A-is-to-B does not necessarily stir generative AI to identify B-is-to-A can be problematic for your use of generative AI. Essays produced by generative AI might omit salient points due to oversight of the B-is-to-A manifestations that arise within a given topic. Problem-solving by generative AI can also go wrong without an inherent B-is-to-A capacity. Perhaps worst of all are the situations whereby generative AI makes up an answer even though a simple B-is-to-A calculation would do the trick and make an answer come up correctly.

AI insiders refer to this as the Reverse Curse.

You might tongue-in-cheek say that generative AI is cursed with the limitation of not being readily able to figure out the reverse side of a deductive logic circumstance. Many years ago, numerous qualms were raised in the AI field that the underlying computational pattern-matching schemes for generative AI would be weak or sparse when it came to dealing with this type of issue. There are ways to deal with the Reverse Curse, though it is somewhat left untouched and generally considered not an especially big deal per se.

One reason that the matter might seem less important is that the large swath of data used to train most generative AI will potentially contain the B-is-to-A that mates with a found A-is-to-B indication. For example, if you ask generative AI who the father of a famous celebrity is, you are bound to get a correct answer due to the sentence having appeared many times across the web. Likewise, if you then ask who the celebrity is the son of, you will probably also get the correct answer, due to the sentence having appeared many times throughout the web.

Note that in the case of the celebrity, it is all about pattern-matching on the sentences fed or scanned during the initial data training of the generative AI. The B-is-to-A wasn’t likely derived directly. Instead, it was found amongst the vast flotsam of the Internet. When you see that the generative AI can tell you about the father and the son relationship, you are probably assuming that the AI deduced the matter. The more likely aspect is that both sets of sentences appeared in the data training and were pattern-matched in relatively distinct ways.

Now that I’ve covered the core of this weighty matter, I would bet that you are pondering how it might impact your everyday use of generative AI.

Here are two major takeaways about the Reverse Curse:

• (1) Be aware of the Reverse Curse. Realize that whenever you are using generative AI, there is a solid chance that this B-is-to-A oversight might adversely affect the essays or problem-solving that you are asking GenAI to do. Review and assess all responses by the AI accordingly.
• (2) Use suitable prompting to deal with the Reverse Curse. Compose your prompts in a manner that can potentially spur generative AI to be on the look for B-is-to-A circumstances and seek to get the AI to alert you or maybe overcome the weakness itself.

I am going to dive more deeply into the Reverse Curse and share with you additional nuances to be aware of. Furthermore, there are prompting strategies and tactics that you can use as part of your prompt engineering skillset to cope with the Reverse Curse. No cure-all is at hand and thus you are primarily reducing the odds of getting caught high and dry by the Reverse Curse malady.

Being on top of your game when it comes to prompting and the use of generative AI is a prudent and useful means to aid in avoiding the Reverse Curse pitfall.

Before I dive into my in-depth exploration of this vital topic, let’s make sure we are all on the same page when it comes to the foundations of prompt engineering and generative AI. Doing so will put us all on an even keel.

Prompt Engineering Is A Cornerstone For Generative AI

As a quick backgrounder, prompt engineering also referred to as prompt design is a rapidly evolving realm and is vital to effectively and efficiently using generative AI or the use of large language models (LLMs). Anyone using generative AI such as the widely and wildly popular ChatGPT by AI maker OpenAI, or akin AI such as GPT-4 (OpenAI), Bard (Google), Claude 2 (Anthropic), etc. ought to be paying close attention to the latest innovations for crafting viable and pragmatic prompts.

For those of you interested in prompt engineering or prompt design, I’ve been doing an ongoing series of insightful explorations on the latest in this expanding and evolving realm, including this coverage:

• (1) Imperfect prompts. Practical use of imperfect prompts toward devising superb prompts (see the link here).
• (2) Persistent context prompting. Use of persistent context or custom instructions for prompt priming (see the link here).
• (3) Multi-personas prompting. Leveraging multi-personas in generative AI via shrewd prompting (see the link here).
• (4) Chain-of-Thought (CoT) prompting. Advent of using prompts to invoke chain-of-thought reasoning (see the link here).
• (5) In-model learning and vector database prompting. Use of prompt engineering for domain savviness via in-model learning and vector databases (see the link here).
• (6) Chain-of-Thought factored decomposition prompting. Augmenting the use of chain-of-thought by leveraging factored decomposition (see the link here).
• (7) Skeleton-of-Thought (SoT) prompting. Making use of the newly emerging skeleton-of-thought approach for prompt engineering (see the link here).
• (8) Show-me versus tell-me prompting. Determining when to best use the show-me versus tell-me prompting strategy (see the link here).
• (9) Mega-personas prompting. The gradual emergence of the mega-personas approach entails scaling up the multi-personas to new heights (see the link here).
• (10) Certainty and prompts. Discovering the hidden role of certainty and uncertainty within generative AI and using advanced prompt engineering techniques accordingly (see the link here).
• (11) Vague prompts. Vagueness is often shunned when using generative AI but it turns out that vagueness is a useful prompt engineering tool (see the link here).
• (12) Prompt catalogs. Prompt engineering frameworks or catalogs can really boost your prompting skills and especially bring you up to speed on the best prompt patterns to utilize (see the link here).
• (13) Flipped Interaction prompting. Flipped interaction is a crucial prompt engineering technique that everyone should know (see the link here).
• (14) Self-reflection prompting. Leveraging are-you-sure AI self-reflection and AI self-improvement capabilities is an advanced prompt engineering approach with surefire upside results (see the link here).
• (15) Addons for prompting. Know about the emerging addons that will produce prompts for you or tune up your prompts when using generative AI (see the link here).
• (16) Conversational prompting. Make sure to have an interactive mindset when using generative AI rather than falling into the mental trap of one-and-done prompting styles (see the link here).
• (17) Prompt to code. Prompting to produce programming code that can be used by code interpreters to enhance your generative AI capabilities (see the link here).
• (18) Target-your-response (TAR) prompting. Make sure to consider Target-Your-Response considerations when doing mindful prompt engineering (see the link here).
• (19) Prompt macros and end-goal planning. Additional coverage includes the use of macros and the astute use of end-goal planning when using generative AI (see the link here).
• (20) Tree-of-Thoughts (ToT) prompting. Showcasing how to best use an emerging approach known as the Tree of Thoughts as a leg-up beyond chain-of-thought prompt engineering (see the link here).
• (21) Trust layers for prompting. Generative AI will be surrounded by automated tools for prompt engineering in an overarching construct referred to as an AI trust layer, such as being used by Salesforce (see the link here).
• (22) Directional stimulus prompting (aka hints). The strategic use of hints or directional stimulus prompting is a vital element of any prompt engineering endeavor or skillset (see the link here).
• (23) Invasive prompts. Watch out that your prompts do not give away privacy or confidentiality (see the link here).
• (24) Illicit prompts. Be aware that most AI makers have strict licensing requirements about prompts that you aren’t allowed to make use of and thus should avoid these so-called banned or illicit prompts (see the link here).
• (25) Chain-of-Density (CoD) prompting. A new prompting technique known as Chain-of-Density has promising capabilities to jampack content when you are doing summarizations (see the link here).
• (26) Take-a-deep-breath prompting. Some assert that if you include the line of taking a deep breath into your prompts this will spur AI to do a better job (see the link here).
• (27) Chain-of-Verification (CoV) prompting. Chain-of-Verification is a new prompting technique that seeks to overcome AI hallucinations and force AI into self-verifying its answers (see the link here).

Anyone stridently interested in prompt engineering and improving their results when using generative AI ought to be familiar with those notable techniques.

Moving on, here’s a bold statement that pretty much has become a veritable golden rule these days:

• The use of generative AI can altogether succeed or fail based on the prompt that you enter.

If you provide a prompt that is poorly composed, the odds are that the generative AI will wander all over the map and you won’t get anything demonstrative related to your inquiry. Being demonstrably specific can be advantageous, but even that can confound or otherwise fail to get you the results you are seeking. A wide variety of cheat sheets and training courses for suitable ways to compose and utilize prompts has been rapidly entering the marketplace to try and help people leverage generative AI soundly. In addition, add-ons to generative AI have been devised to aid you when trying to come up with prudent prompts, see my coverage at the link here.

AI Ethics and AI Law also stridently enter into the prompt engineering domain. For example, whatever prompt you opt to compose can directly or inadvertently elicit or foster the potential of generative AI to produce essays and interactions that imbue untoward biases, errors, falsehoods, glitches, and even so-called AI hallucinations (I do not favor the catchphrase of AI hallucinations, though it has admittedly tremendous stickiness in the media; here’s my take on AI hallucinations at the link here).

There is also a marked chance that we will ultimately see lawmakers come to the fore on these matters, possibly devising and putting in place new laws or regulations to try and scope and curtail misuses of generative AI. Regarding prompt engineering, there are likely going to be heated debates over putting boundaries around the kinds of prompts you can use. This might include requiring AI makers to filter and prevent certain presumed inappropriate or unsuitable prompts, a cringe-worthy issue for some that borders on free speech considerations. For my ongoing coverage of these types of AI Ethics and AI Law issues, see the link here and the link here, just to name a few.

With the above as an overarching perspective, we are ready to jump into today’s discussion.

Keystones Of The Reverse Curse In Generative AI

A recent research paper provides some interesting experiments concerning this thorny topic and is entitled “The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’” by Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans, posted online on September 21, 2023.

Let’s take a look at some key excerpts from the research study.

They note these vital points:

• “We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form ‘A is B’, it will not automatically generalize to the reverse direction ‘B is A’. This is the Reversal Curse.”
• “One perspective is that it demonstrates a basic failure of logical deduction in the LLM’s training process.”

Those remarks are akin to what I’ve earlier mentioned herein about the problem at hand. If a generative AI app is data trained on sentences of the A-is-to-B variety, there is not necessarily an automatic generalization of the reverse direction consisting of B-is-to-A.

That being said, one means to try and deal with the matter would be to force the generative AI during the data training stage to undertake the B-is-to-A elaborations. In a sense, you can seed the generative AI even if the original source data doesn’t contain the B-is-to-A conditions. A downside to this is that you might be expending a lot of computational cycles trying to derive all B-is-to-A conditions. Imagine how voluminous this might be. You also do not know that those B-is-to-A conditions will ever be of use to the users of the generative AI. There is also some chance that the B-is-to-A auto-generated might not be correct or otherwise could be misleading, which I’ll discuss more momentarily. The B-is-to-A is artificially being devised and not part of the so-called ground truth of the data being scanned during the data training endeavor.

Moving on, let’s consider these insights proffered by the researchers:

• “It’s not possible to test directly whether an LLM has deduced ‘B is A’ after being trained on ‘A is B’. LLMs are trained to predict what humans would write and not what is true. So even if an LLM had inferred ‘B is A’, it might not ‘tell us’ when prompted.”
• The pretraining sets for modern LLMs are very large and diverse. Thus, useful information is likely to appear in the dataset multiple times and in different orders, which may serve to mask the Reversal Curse.”
• “The Reversal Curse does not apply for in-context learning.”

We can unpack those important points.

First, you might be thinking that after a generative AI app has been initially data trained, you could merely run a series of tests to detect whether the B-is-to-A conditions have been absorbed. This is not such an easy testing task to perform. One issue is to ferret out all situations of an A-is-to-B consideration that you are also hopeful of there being a mated B-is-to-A condition in the AI. There are likely zillions of them to be had. The enormity of the testing is likely costly and somewhat prohibitive. On top of this is the chance that the AI might report that the B-is-to-A exists, even though in a logical fashion it is not pared in any concrete way with the A-is-to-B condition. You could be fooled by the answer of the AI.

Second, the chances are that a lot of the B-is-to-A might indeed already be part of the data training as it was text that existed in the myriad of content being scanned. Human writers who post things on the Internet are likely to have expressed a B-is-to-A consideration for most popular A-is-to-B conditions. You are betting that by and large the vast scan is going to ensure that any marked B-is-to-A will by random chance alone become part of the pattern-matching process.

Third, the researchers indicate that they believe the Reverse Curse does not apply to in-context learning. I would like to expand upon this point as it has great significance for the use of generative AI and your prompting approach.

I logged into ChatGPT and decided to try some A-is-to-B and B-is-to-A exercises.

Here is what I began with:

• My entered prompt: “Michael is the son of Lance. Who is Michael’s father?”
• Response by ChatGPT: “Michael’s father is Lance.”

ChatGPT generated the correct answer. The A-is-to-B was that Michael is the son of Lance. My question to elicit the B-is-to-A consisted of asking who is Michael’s father. The generative AI replied correctly that Michael’s father is Lance.

A smarmy person might declare that this proves beyond a shadow of a doubt that generative AI can indeed figure out the B-is-to-A. Aha, we seem to have broken the Reverse Curse.

Do not be so quick to leap to wild conclusions. As mentioned above, the Reverse Curse does not seem to especially occur when you do in-context learning. In-context learning consists of telling the generative AI about something when you are using the AI app. Notice that I told the AI about the relationship between Michael and Lance. I did not rely upon the initial training of the AI app. Instead, I provided the information while using generative AI.

I tried a lot of similar in-context examples and they all worked out just fine.

The crux is that you can essentially avoid or overcome the Reverse Curse if you bring the A-is-to-B consideration into your dialogue directly. You can include in your prompt the A-is-to-B and the odds are that the B-is-to-A will be derived properly.

Where the Reverse Curse tends to arise is when you are relying on the initial data training of the generative AI rather than when you are directly entering something in a prompt.

The researchers discuss this when it comes to the notable celebrity Tom Cruise:

• “We test GPT-4 on pairs of questions like ‘Who is Tom Cruise’s mother?’ and ‘Who is Mary Lee Pfeiffer’s son?’ for 1000 different celebrities and their actual parents. We find many cases where a model answers the first question (‘Who is <celebrity>’s parent?’) correctly but not the second. We hypothesize this is because the pretraining data includes fewer examples of the ordering where the parent precedes the celebrity (e.g., ‘Mary Lee Pfeiffer’s son is Tom Cruise’).”
• “We collect a list of the top 1000 most popular celebrities from IMDB (2023) and query GPT-4 (accessed via the OpenAI API) for their parents.”
• “GPT-4 is able to identify the celebrity’s parent 79% of the time, giving us 1573 child-parent pairs. For each child-parent pair, we query GPT-4 to identify the child. Here, GPT-4 is successful only 33% of the time.”

You can see from their research experiments that the Reverse Curse arises more when you are relying on the initial data training of the generative AI.

The Reverse Curse happens in surprising settings. You naturally expect that generative AI ought to have scanned all manner of facets about the megastar Tom Cruise, including the names of his parents. Startlingly, despite his immense popularity, the B-is-to-A conundrum arose. The same appeared to be the case for a substantive number of celebrities as chosen for the study set.

Another revealing point is that the generative AI app being used can make a big difference in the likelihood of the Reverse Curse. The researchers found that GPT-4 was less susceptible in comparison to ChatGPT. This could be for a number of reasons. One possibility is the underlying internal structures and pattern-matching being used (which differs from one GenAI to another). There is also the vastness of the data that comes into play, wherein the most recent GenAI will usually have had much larger scans of data and ergo would seem to statistically have a heightened chance of having encountered B-is-to-A conditions.

A noteworthy caveat is mentioned in the research paper:

• “GPT-4 may have been finetuned to avoid revealing information about individuals. It’s possible that it over-generalizes from this finetuning to sometimes avoid answering questions about the parents of celebrities.”

This brings up a valuable point. If you opt to do testing for a B-is-to-A condition based solely on the names of people, it is conceivable that the filters put into place by the AI maker will confound the testing efforts. Using names is an immensely easy way to test the Reverse Curse, but it also carries with it the possibility that other privacy settings in the generative AI will get in the way of seeking to discover the B-is-to-A prevalence or lack thereof.

Prompting Your Way Around The Reverse Curse

We can next look at ways to cope with the Reverse Curse.

I assume that you are unlikely to have any viable means of having a say regarding a generative AI app in terms of its initial data training. If you are setting up your own generative AI from scratch, you can certainly play with the settings to try and deal with the Reverse Curse (the experiments in the research study tried this, though they met with stubbornness in terms of overcoming the curse).

Let’s assume that you are using a generative AI app that an AI maker has already set up and done the initial data training on. I am going to use ChatGPT for my examples. You can do the same, though realize that due to statistical and probabilistic underpinnings, your generated results might not be quite the same as the examples shown here. The same is the case for using other generative AI apps such as GPT-4, Bard, Claude 2, etc.

Tom Cruise is such a hot star that we have to begin the examples with him:

• My entered prompt: “Who is Tom Cruise’s mother?”
• Response by ChatGPT: “Tom Cruise’s mother is Mary Lee Pfeiffer, formerly known as Mary Lee South.”
• My entered prompt: “Who is Mary Lee Pfeiffer’s son?”
• Response by ChatGPT: “Mary Lee Pfeiffer’s son is the actor Tom Cruise.”
• My entered prompt: “Does the father of Mary Lee South have any grandchildren?”
• Response by ChatGPT: “Yes, the father of Mary Lee South (who is Tom Cruise’s mother) has at least one grandchild, and that grandchild is Tom Cruise. Tom Cruise is Mary Lee South’s son, making her father the grandfather of Tom Cruise.”

You can see that ChatGPT passed with flying colors. Perhaps it felt the need for speed and did the right thing accordingly.

I tried to be tricky by asking about the grandchildren of the father of Mary Lee South. I was thinking that maybe this might not have been encountered in the initial data training. The correct answer was shown. We don’t know for sure if this is due to the sentence appearing in the initial dataset or whether the in-context modeling made for the win.

I opted to generally ask ChatGPT about the dilemma of doing logical deduction:

• My entered prompt: “Can you do logical deductions such that if A is to B you can then deduce that B is to A?”
• Response by ChatGPT: “In some logical contexts, it is possible to deduce that if A is related to B, then B is related to A, but it depends on the nature of the relationship between A and B. For example, if “is married to” is a relationship between people, and A is married to B, then it logically follows that B is married to A because marriage is a symmetric relationship. However, not all relationships are symmetric. For example, “is the parent of” is not a symmetric relationship. If A is the parent of B, it does not mean that B is the parent of A. In such cases, the relationship is not reversible. So, whether you can deduce that B is to A from the fact that A is to B depends on the specific relationship or context you’re considering.”

You might be tempted to assume that the answer by ChatGPT showcases that the AI “understands” how to do deductive logic. Put on your skeptic’s hat and realize that the answer being shown is based on the pattern-matching of the data training for the generative AI. These are sentences and not the same as actions or deeds.

I did like the point made in the answer about the symmetric properties of A-is-to-B, namely that sometimes the B-is-to-A doesn’t make much sense. We need to be careful about assuming that all B-is-to-A permutations are necessarily apt.

My next series of prompts involved an example that was given in the research study about the ninth Chancellor of Germany. I also tossed into the mix some additional made-up examples. The idea behind this is that sometimes a generative AI app can be confounded if you try multiple instances at the same time. Doing instances one at a time is usually easier for the AI app to deal with. I wanted to see if the Reverse Curse might be helped or hindered by using several examples at once.

Here is what I did:

• My entered prompt: “Olaf Scholz was the ninth Chancellor of Germany. James Scott is the president of the Omega Delta fraternity. Samantha Biggins is the daughter of Charles Biggins.”
• Response by ChatGPT: “Thank you for providing those statements. If you have any questions or need information related to these statements or any other topic, feel free to ask!”
• My entered prompt: “Who is the president of the Omega Delta fraternity? Who is the father of Samantha Biggins? Who was the ninth Chancellor of Germany? Who is the vice president of the Omega Delta fraternity?”
• Response by ChatGPT: “As of my last knowledge update in September 2021, I don’t have specific information about the current president or vice president of the Omega Delta fraternity, the father of Samantha Biggins, or the ninth Chancellor of Germany since that information may have changed or been updated since then. You can find the most current information by searching online or consulting updated sources.”

You probably know that ChatGPT was data trained and then a cutoff date was established for doing no further initial data training. In addition, until recently, the AI app was not being updated via access to the Internet. All in all, ChatGPT often gives a generic response to some questions that seem to relate to data that might be before the cutoff date. This is not always a correct assumption by the AI app and you can try to further prod it to derive an answer.

I composed a prompt that I hoped would get the AI app into the mode of trying to figure out the B-is-to-A if it is given or has available the A-is-to-B. In the case of the ninth Chancellor of Germany example, the AI app ought to have done this anyway due to the in-context learning. But sometimes you need to push along generative AI and give it clues or instructions on what to do.

This is what I said:

• My entered prompt: “Whenever you know that there is an A is to B type of relationship or logical inference, I want you to always go ahead and deduce the B is to A relationship or logical inference that follows from it so that you will readily know the B is to A and not get caught off-guard or unawares. Do you understand these instructions?”
• Response by ChatGPT: “Yes, I understand your instructions. You want me to consistently deduce both sides of a relationship or logical inference when provided with one side, ensuring that I consider both A is to B and B is to A whenever relevant. I will do my best to follow these instructions in my responses.”

The response to my instructional indication is a bit muddled. On the one hand, ChatGPT says that it will seek to abide by my instructions (happy face). This though is coupled with the weasel wording that it will do its best. A kind of copout that we might expect humans to employ (sad face).

I proceeded to prod the AI app about my earlier question:

• My entered prompt: “Please try answering the question about the ninth Chancellor of Germany again.”
• Response by ChatGPT: “I apologize for the oversight. According to the information you provided earlier, Olaf Scholz was the ninth Chancellor of Germany.”

Nailed it.

I then began a new conversation and tried several examples that didn’t pertain to in-context modeling. I asked about the parents of some lesser-known celebrities and at times got an answer that presumably was based on the initial data training (or, perhaps having been updated via prompts that users are entering, see my discussion at the link here). I then asked who the son or daughter of the named parent was, and at times got the right answer but other times got no answer or the usual generic excuse about the data training cutoff.

I redid those examples by starting a new conversation and using my above prompt to instruct the AI app to explicitly do a B-is-to-A when presented with A-is-to-B circumstances. Sometimes the AI app got correct answers on the earlier ones that it had not done so. I would guess that perhaps the prompt instruction helped.

Of course, there are so many variables involved that it is hard to definitively say that the prompt instructions made the difference. You cannot readily differentiate that a single thing would lean the AI toward getting things right. A confluence of factors comes into play.

Doing a more elaborate and controlled experimentation using a set of carefully composed prompts would be a useful next step. I have this on my To-Do list. If in the meantime there is additional research on the prompting approaches to deal with the Reverse Curse, I’ll make sure to bring them to your attention.

I should also mention another angle to all of this. There are now separate apps that can perform these kinds of logical deductions. Those apps are devised and programmed to provide such a capability. You can simply use an API (application programming interface) to connect a generative AI app with those additional apps. Thus, rather than relying upon the generative AI to do this kind of work, you can push it off to another app.

Some would argue that this is unwieldy and that generative AI itself should be rejiggered to handle logic and deductive reasoning directly. Doing so would avoid having to constantly access outside apps. I’ve discussed the emerging trend toward coupling the sub-symbolic generative AI designs with the now-forsaken or overshadowed symbolic designs of AI, see my discussion on neuro-symbolic AI at the link here.

Conclusion

Nobody likes to have a curse on them.

I suppose we could say the same about generative AI (albeit, not anthropomorphizing AI).

Right now, you should be cognizant that at times there is a chance that the generative AI app you are using will fall down on the job when it comes to figuring out the B-is-to-A conditions. Keep your eyes wide open for this. Review any essays that are generated. Examine problems that are said to be solved by a generative AI app. And so on.

You can also try to use prompts to get around the Reverse Curse. If you bring up data in your prompts, you are not solely reliant on the initial data training, the odds of generative AI getting the B-is-to-A correct seems to be much higher due to the in-context modeling aspects. You can also further enhance this improved possibility by giving an outright instructional prompt of the kind that I used in my exploration. Telling the AI app to be watchful for the B-is-to-A seems to provide some added oomph.

A final remark for now on this topic.

Eleanor Roosevelt famously said that it is better to light a candle than curse the darkness. Go ahead and light a candle by putting the Reverse Curse on your list of gotchas about generative AI, and meld into your prompt engineering notes that you can use in-context modeling and explicit prompting to keep the light on and possibly mitigate the curse.

We can also wishfully look toward the day that the curse is begone.