We recently encouraged writers (of all kinds) in our recent Insider’s Insight to experiment with large language artificial intelligence (AI) models when drafting their documents [1]. The guide provides instruction on the safe use of these AI tools and recommends that only through practice can you generate valued outputs. In the guide we discussed various approaches and techniques, but we didn’t provide worked examples. To highlight some of the limitations you can meet with AI and to counter our earlier omission of a worked example, I have attempted to provide a timely worked example.

In recognition of this being Valentine’s Day I have used ChatGPT to address the perennial challenge for non-romantics – how to write something meaningful in my partners Valentine’s Day card. Valentine’s Day feels like a particularly suitable topic for anyone in the pharmaceutical industry exploring how AI can best support their writing. We equate Valentine’s Day with love, a biological function (tenuous link) and one scientists are often a little too detached from to create heartfelt missives. Furthermore, any business-related question you might want to investigate poses an organisational risk [2].

To minimise risk and track your interactions we recommend in our Insider’s Insight that the full details of any interactions with AI are recorded and that your proposed investigation undergoes a thorough risk assessment and obtains prior managerial approval (assuming you are using company equipment) [1]. In the present exercise the AI tool used was ChatGPT3, the defined purpose of the exercise fell under both ‘first draft’ and ‘brainstorming’ (both permitted use at Niche [1]) and the associated organisational risk was rated as ‘low.’ The interaction was performed on 7 Feb, 2023 and it took the programme just under 45 seconds to produce a 934-word response. The request submitted to ChatGPT3 was:

“Write a 1000-word scientific essay on love in the style of Marcus Aurelius.”

I chose the question to combine a quintessential aspect of the human condition (love) while also having the AI to consider the scientific aspects of love. As AI's can generate responses from different viewpoints, I secretly hoped the exercise might generate a (brief) poetic masterpiece that I could use in this year’s obligatory Valentine’s Day card (I believe that currently an AI cannot sue you for plagiarism in a Valentine’s card). Below, I have broken down the response to see if we can learn anything about ChatGPT3’s capabilities and limitations.

The AI selected the above title and it sounds sexy. The essay opened with a cool, Brian Cox-esq reference to the cosmos but I am not sure the average reader is going to relate to the word “behooves.” I am also not a great fan of the AI’s over-explaining of the intent behind its essay in the introduction (it gives away the plot). However, I did enjoy the use of “contemplative journey” and “intricate tapestry.” In short, the opening felt robust, though a quick Google search shows that these two concepts are not unique.

In creating its essay, ChatGPT split the body of its article into five key areas, the first being “The Nature of Love.” I was a slightly disappointed that the first sentence for this section focused on explaining how Marcus Aurelius would be expected to respond. I feel that a human writer would take that as implicit.

The algorithm subsequently contemplates on how love, whether romantic, familial or platonic, is not an external force but an internal response to our perceptions. As a scientist I had hoped for a little more reference to the impact of love on the central nervous system. It is possible that my instruction failed to consider that in Marcus’ time science itself wasn’t a ‘thing.’ English philosopher and historian of science William Whewell coined the term scientist in 1833, and it first appeared in print in Whewell's anonymous 1834 review of Mary Somerville's On the Connexion of the Physical Sciences published in the Quarterly Review. ChatGPT’s narrative continues to consider wisdom and indifference to love, which feels a little unromantic (considering the date) and then mentions how the Stoic seeks balance and temperance. It was a pity that ChatGPT missed another opportunity to bring in some details on how science has identified romantic love as having a biochemical signature similar to that of addiction and not at all indifferent [3].

Moving on to virtue, ChatGPT underlined how love, when guided by wisdom, enhances our understanding of others and fosters empathy. However, my own experience of love (and I expect that of most of us) is that it is beset with confusion, counter-messages and uncertainty. What about the suffering? Even requited love can be painful – isn’t that why they call it lovesickness? Most of us can describe the causes, pathogenesis and cure. The disorder has been attested to in the medical literature since classical times and may still have a place in current medicine in the frame of psychiatry and humoral disorders. Although in different cultures there is a general agreement on the symptoms, including fever, agitation, loss of appetite, headache, rapid breathing, and palpitations, the treatments vary greatly in the various cultural contexts [4]. I was left feeling that the AI had missed the underlying humanity of everyday attachments in focusing on the nature of ‘perfect’ love.

ChatGPT next reflected on our vulnerability and selfishness – certainly human qualities – and on the necessity for courage and justice. The age-old philosophical, biological, and social debate over the basic nature of humans as being ‘universally selfish’ or ‘universally good’ continues today highlighting sharply divergent views of natural social order. Yet again, I felt that the AI missed an opportunity to adopt a more biochemical theme that could bind together all aspects of our personalities, perhaps referring to the soup of pharmacologically active neurotransmitters that wash across our hapless brains. Observing how the underlying neural circuitry differs between psychopaths and altruists with emotional processing being profoundly muted in psychopaths and significantly enhanced in altruists would have caught my attention [5].

Should we be surprised that an AI is ultimately literal? In this case ChatGPT seems to have conveniently overlooked that grief is ultimately the flip side of love. The essay cautions against attachment and dependence, referencing its transient nature but misses the emotional consequences of loss. And again, no mention of science, I would have been delighted to have read about how, in the future, drugs may be able to fix our romantic lives [6]. But this is the joy of ChatGPT, I could simply modify my prompt and in less than a minute have a new essay without these limitations.

In the words of Savage Garden’s global top-40 hit, ‘Affirmation’ (2000), “I believe that you can’t appreciate real love until you’ve been burned.” It beggars the question, were the lyrics of virtually half the pop songs ever written missed off ChatGPT’s recommended reading list? Love, heartbreak, and relationships are popular topics for songs because they are universal human experiences that most people can relate to. Today’s pop songs are full of problematic messages when it comes to romantic love and relational fidelity. They tend to focus on either the ‘hottest’ or ‘darkest’ elements of short-term relationships. ChatGPT however has stayed laser-focused on the requested style. If you want to capture a broader understanding you need to make this clear in the challenge you make to your AI.

Clearly, the initial request limited the AI’s output to what would have been available to an author in the times of Marcus Aurelius (circa 160 AD) – thus confounding any attempt to include modern-day science. It highlights importance of how your pose your request. Nevertheless, it also brings to the fore the power of AI in giving you the opportunity to simply repeat the exercise.

The fourth paragraph focuses on time. What an opportunity for another Brian Cox-like comment on the unimaginable scale of space and time and the impermanence of love – confirmed by every Country & Western song (ever!). The response however reflects more of an appreciation of the underlying philosophy of love practiced by the perfect individual with no scientific consideration.

In the penultimate paragraph, ChatGPT considers universal nature of love, extending relationships to encompass a broader perspective. This raises an issue often faced by AIs based on large language models, that of the source material. AIs can only generate answers from the information they have access to (or the hallucinations they make up themselves). An important consideration is that any bias contained in the source data will be propagated by the AI. Obviously, certain sections of our collective knowledge is locked behind paywalls. For example, many contemporary scientific publications, journals and newspaper stories are pay-to-view. And yet, we are also seeing more and more historic publications becoming freely available on the internet. ChatGPT’s outputs can only reflect the opinions and biases of the data available to it – putting the case crudely, what would AI make of porn? According to various sources, around 35% of all internet traffic is generated by porn sites. In this sense, ChatGPT is clearly unaware of our more lascivious natures.

As might be expected, ChatGPT closes its exposition with a conclusion. One observation I often make when reviewing the closing arguments of less experienced writers I find that rather than extrapolating on their report, they tend to repeat observations that have already been discussed. It takes a certain level of understanding, vision and confidence for a writer to build on their findings. This skill develops as writers mature. I shouldn’t therefore be too surprised that in its concluding statement, like any other youthful author, ChatGPT chose to simply explore alternative ways to re-express its earlier thoughts. I would note however, that it does this very eloquently. The AI clearly has a vocabulary larger than many young writers and is not afraid to use it.

Despite its limitation, experimenting with AI highlights one of the fundamental lessons any budding scientist eventually comes to appreciate – garbage in, garbage out (GIGO). The spirit of GIGO was first expressed (in print at least) in 1962 by George Fuechsel, an IBM programmer. He succinctly observed that: if we put poor information into our computer models, we will get poor information out of them. Expect controversy, bad insights, poor decisions, and bad policy to follow. Clearly, the art to using AI is in setting intelligent prompts and practice makes perfect.

In closing, it seems that those who speculate on the ‘abilities’ of AI suggest that it is soulless, that it simply seeks to regurgitate mathematically modelled mediocrity. However, the closing poetical sentence of the current exercise seems to counter the argument. I am not sure I would be unhappy if I had written:

“As we reflect on love through the wisdom of Marcus Aurelius, we are reminded that in the vastness of the cosmos, love remains a beacon of light, guiding us toward a virtuous and meaningful existence.”

Now, where did I put that Valentine’s Day card?

References

1. Artificial Intelligence in Medical Writing: An Insider’s Insight. https://www.niche.org.uk/asset/insider-insight/Insiders%20Insight%20GAILs.pdf
2. Hardman TC. Love Pharmacology. https://www.linkedin.com/pulse/love-pharmacology-tim-hardman/?trackingId=IVxJ%2BsDPRNCfWXN4N09%2Bkw%3D%3D
3. Earp BD, Wudarczyk OA, Foddy B, Savulescu J. Addicted to love: What is love addiction and when should it be treated? Philos Psychiatr Psychol. 2017 Mar;24(1):77-92.
4. Lippi D. Arrhythmias in the History: Lovesickness. Card Electrophysiol Clin. 2017 Sep;9(3):341-344.
5. Sonne JWH, Gash DM. Psychopathy to Altruism: Neurobiology of the Selfish-Selfless Spectrum. Front Psychol. 2018 Apr 19;9:575.
6. Earp BD and Savulescu J. Love is the Drug: The chemical future of our relationships. Manchester University Press, 2020 ISBN: 9781526145413