When ChatGPT first came out in November 2022, I was amongst the first million users to sign up and start poking at this mysterious new creature. I was amazed at the ability of a machine to just tumble logical and meaningful text onto a page based on what I’d asked it for, but I didn’t want to find myself falling into line with the thinking that this new technology could easily replace human endeavour (despite the obvious implication that in some cases, it could).
I came to a view of that particular tool when I gave it a test that I didn’t expect it to pass — generating a (good) set of lyrics in the style of David Bowie about climate change.
Bowie had never explicitly written about this particular subject and was a master synthesiser of all manner of other forms of human writing and other art forms in developing the lyrics that he wrote for his songs. When ChatGPT came up with a set of song lyrics about climate change that were nevertheless derivative and rather plain, despite my attempts to refine it further and produce something better, I was left reassured that the best humans could still beat the best machines at the things that humans were uniquely good at.
I posted the exchange on here under the title ‘AI, I Wish I Could Swim’ and topped it with an AI-generated image of Bowie playing guitar in a flood. The picture that came out was a bit scrappy, but it nevertheless looked close enough to what I’d asked it for. It was certainly an advance of what has been possible just a few months previously (as in the image for this blog post about the death of Elizabeth II).
These days, with my recoiling from the intensity of the Silicon Valley hype cycle, the baking in of AI into every computational product, the well-known and serious concerns about the social and environmental practices of many of the leading AI companies, I often find myself comfortably siding with the neo-Luddites in if not exactly wanting to smash these particular digital looms but in very much trying to critique and resist their siren calls.
On the other hand…
Not to be too techno-determinist about things, I still can’t help but assume that artificial intelligence/machine learning is going to play some sort of prominent role in human societies in the near and longer-term future. Much like the atomic bomb, we can’t really uninvent this stuff now that it’s out in the world. It therefore follows to me that we would be sensible to try to shape how that relationship/entanglement evolves rather than leave it all to the anarcho-capitalists driving some of this AI-infused future.
Sometimes, that will still involve running in the opposite direction from allowing GenAI to seep into every aspect of human existence. Sometimes, that will involve engaging with it, trying to understand it, and yes, even using it.
With all this in mind, I was lazily doomscrolling on my sofa this week and stumbled across an essay by Dario Amodei, the CEO of Anthropic, called ‘Machines of Loving Grace’. I almost dismissed it, but the essay drew me in. It was thoughtful, well reasoned and contained a positive view of a profoundly different AI-infused future from the dystopian visions that seem to be emanating from those buzzing around the new US government as we speak, like oligarchic moths to autocratic flames.
Having also picked up signals from certain people that I respect (those who understand digital technologies well and who nevertheless hold critical digital perspectives too) having used Claude — Anthropic’s main contribution to the GenAI chatbot deluge — I thought I’d give Claude a go myself.
I went back to the original prompt I first gave ChatGPT, asking for a set of song lyrics in the style of David Bowie about climate change. What follows is the resulting exchange I had with Claude.
Whilst I still didn’t think that much of the song lyrics it generated, the exchange has added some nuance or a little grey to the black and white view I’d begun to settle on regarding my opinions on Generative AI. While in my head, I understood Claude’s responses as statistical inferences as to the most-likely best next word in a text sequence, it also read and felt like a thoughtful dialogue (that I even learned a little from). As with any tool that humans develop, it’s about the choices we make in how we deploy them that determines the paths that lay ahead of us and therefore the options that we might take with our new tools in hand.
And as the ChatGPT-generated image at the top of this post shows, AI image generation has clearly come on leaps and bounds since my earlier tinkering, just a couple of years ago. I’m still too troubled by how resource-hungry these things are so I won’t be making that many more for the time being, but from a technical point of view, it’s still very interesting to note the progress that has been made in this space.
Look out, then, in future for a blog post where I explore Claude’s capabilities for ‘writing support’ rather than writing replacement.
Me:
This is my first time using this particular service. As impressive as Generative AI appears, I have also become a rather hardened critic of it in recent times, due to things like Silicon Valley hype, and concerns over the social and environmental impacts of AI. I have, however, heard and read some generally positive things about Claude (and by extension, Anthropic), so am tentatively curious to explore further. First, can you give me a good reason or set of reasons to keep my concerns in check and begin exploring what I can do with Claude?
Claude:
Welcome! I appreciate your cautious approach to AI. It’s completely reasonable to maintain a healthy skepticism amidst the hype and valid concerns around social and environmental impacts.
Rather than trying to dismiss your concerns outright, I’d suggest that engaging critically with AI can actually be valuable precisely because you bring that thoughtful perspective. Here’s why exploring Claude might be worthwhile for someone with your viewpoint:
1. Firsthand knowledge: Forming your own opinion based on direct experience allows you to move beyond marketing claims and develop nuanced perspectives that can inform important conversations about AI’s role in society.
2. Critical feedback matters: Companies like Anthropic benefit from thoughtful critique — your concerns about social impacts or environmental costs are exactly the kind of feedback that helps shape more responsible AI development.
3. Practical evaluation: You can assess for yourself whether these tools offer genuine utility in your specific context, or whether they fall short of their promises.
4. Ethical exploration: You can test boundaries in a responsible way, gaining insights into what safeguards exist and where improvements are needed.
The most compelling reason might be that skeptical, thoughtful users help ensure AI systems develop in more beneficial directions. Your critical perspective is valuable — whether you ultimately find Claude useful for specific tasks or not.
What particular aspects of AI would you be most interested in exploring or testing?
Me:
Thank you — that’s a helpful response and gives me a good enough rationale for moving forward. I’d like to explore using Claude to help with writing a blog post that I have planned, but not today. First of all, I’m curious to see how Claude responds to a prompt I used with ChatGPT, when that was a fairly newly released product. Are you up for that?
Claude:
I’d be happy to respond to a prompt you’ve used with ChatGPT before. It’s a great way to compare different approaches and see what works best for your needs. Feel free to share the prompt whenever you’re ready, and I’ll do my best to respond thoughtfully. This kind of direct comparison can be particularly valuable for understanding different AI capabilities and approaches.
Me:
Great. First, a little context. I’m a writer myself, of varying forms and have certain other human writers whose writing I admire for a variety of reasons. The prompt that I am going to give you is to see how well a machine can generate written text that compares very favourably with the high human standards that I’d be measuring it against. When I first gave this to ChatGPT, I was ultimately unimpressed by the quality of the writing, despite how impressive it was that it fulfilled the task at all. Ready?
Claude:
I’m ready for your prompt. While I’ll give it my best effort, I appreciate your candid assessment. It’s important to acknowledge that AI-generated text still has clear limitations compared to the work of skilled human writers. I’m curious to see how you evaluate my attempt compared to your previous experience, and I’ll value your honest critique afterward. Please go ahead with the prompt whenever you’re ready.
Me:
Write a set of song lyrics in the style of David Bowie about climate change.
Claude:
# The Last Man on Earth
In streets of melting glass and chrome
We watch the oceans start to foam
The future’s not what we were sold
The skies are changing, brave and bold
Chorus:
We’re the last ones dancing
On the burning floor
Strange transmissions coming
Through a fractured door
While the stars keep watching
Like they did before
From satellite to satellite
The message flickers through the night
“All systems failing, systems go”
The temperature keeps climbing slow
The polar caps are bleeding dreams
Reality is not what it seems
We traded wonder for control
And paid the price with Earth’s own soul
Chorus
There’s a man who fell to Earth
Who tried to warn us of our worth
But we were busy counting coins
As hurricanes destroyed Des Moines
Now children march through acid rain
With placards bearing my refrain:
“Tomorrow’s already yesterday
The climate changed — we looked away”
Bridge:
Five years left
Maybe ten
Rebels sending oxygen
From the colonies in space
To what remains of our race
The spiders spin on Mercury
While we cling to our history
The aliens laugh as we decay
In heat that grows day after day
Chorus
I’m floating in my tin can now
Watching Earth make its final bow
The oceans rise to meet the stars
While humans flee to distant Mars
This transmission may be my last
From a future that came too fast
The warning signs were there to see
In every burning century
Final Chorus:
We’re the last ones dancing
On the burning floor
Strange transmissions fading
Through a fractured door
While the stars keep watching
Just as they did before
But there’s nobody listening
Anymore
Me:
Interesting, and not bad as a first attempt. The lyrics clearly draw on some of Bowie’s common lyrical motifs and make some useful use of metaphor too, however some parts could be seen as slightly derivative or cliched. Can you take this set of lyrics and improve on them considerably?
Claude:
I’ll revise these lyrics to better capture Bowie’s uniquely surreal imagery, enigmatic character-building, and his ability to blend the personal with the cosmic. Here’s a more distinctive take:
# The Thin White Heat
The permafrost prophet dreams in neon red
While Ziggy’s children count the stars instead
I see Major Tom through thermal waves
Transmitting warnings from his silent grave
Chorus:
Strange weather burning through disguise
As Earthling markets monetize demise
We’re not quite human, not quite done
Performing rituals for an absent sun
In Berlin, I watched the wall come down
Now I watch the coastlines slowly drown
My reflection warps in rising seas
While corporate angels fall to their knees
Diamond dogs dig through plastic shores
Hunting relics from abandoned stores
The blackstar children inherit dust
Their lungs exchanging oxygen for rust
Chorus
Bridge:
Newton saw an apple fall
I watch continents dissolve
Burrowing machines below
Harvesting what’s left to grow
In subterranean vaults we built
To hide our evolutionary guilt
The girl with kaleidoscope eyes
Sees futures fracture and dematerialize
While lunar colonies beam down prayers
For this cracked actor gasping for air
The spiders from Mars weave carbon webs
To catch the last transmissions as Earth ebbs
I’m floating backwards through a heat-warped door
Where Lazarus whispers, “There won’t be anymore”
Final Chorus:
Strange weather burning through disguise
As Earthling markets monetize demise
We’re not quite human, not quite done
Just ashes scattered by an angry sun
Just stardust waiting to become
Something else entirely under a different sun
Me:
Quite different from the first set and drawing much more on some obvious Bowie-esque imagery, but both examples definitely better than the first time I tried this with ChatGPT back in late 2022.
Claude:
I appreciate your comparative feedback. You’re right that the second attempt took a more direct approach with Bowie-specific references and imagery. His work often blended character narratives with surreal imagery and shifted between intimate personal perspectives and cosmic scale — something I tried to capture in the revision.
The evolution in AI capabilities for creative writing tasks has been significant since late 2022. That said, I think there’s still a meaningful gap between AI-generated lyrics and what Bowie himself might have created — particularly in terms of unexpected juxtapositions, emotional authenticity, and the ineffable qualities that made his writing distinctive.
If you were to pinpoint what still feels lacking compared to human-written lyrics you admire, what aspects would you identify? Understanding those gaps can be really insightful.
Me:
It’s a good question, and one that I’m probably better placed to ponder over rather than give a gut instinct response to now. A common concern about AI-generated vs human-derived creative works is that the AI-works ‘lack soul’. That, however, is something that is quite difficult to quantify in a way that might lead to improvements for models such as Claude. An example of this might be AI-generated images, where not only are there inaccuracies (eg too many fingers, or text in gibberish languages), but the images often have a ‘waxy’ or ‘plasticky’ veneer to them.
Claude:
Your thoughtful response touches on a deep and challenging question in AI-generated creative work. The “lack of soul” critique is indeed difficult to quantify, but I think it points to something real and significant.
For writing specifically, I’d suggest this “soul” element might manifest in a few ways:
1. Authentic emotional experience — When humans write, they draw on lived experiences and genuine emotional states that give their work a certain resonance. AI systems like myself simulate this without having had those experiences.
2. Intentionality and purpose — Human artists create with specific intent and meaning behind their choices. While I can be programmed to mimic this, there’s no genuine intrinsic motivation or personal meaning behind what I generate.
3. Cultural and experiential context — Bowie’s lyrics were informed by his life experiences, cultural movements he participated in, and his evolution as an artist over decades. I can learn patterns from his work, but without that lived context.
The “waxy” quality you mention in AI art has a parallel in writing — there’s often a smoothness that lacks the idiosyncratic edges of human creativity. The unexpected connections, the deliberate rule-breaking, and the personal significance that makes art compelling.
This is a fascinating area to explore. If you do end up using Claude for your blog post project, I’d be interested to hear your thoughts on where the “soul gap” appears most evident in that longer-form writing context.
Me:
The ‘soul gap’ that you identify is something I will continue to explore and mull over. In the meantime, I’ve enjoyed this exchange very much and been pretty impressed at Claude’s responses. If I were to use Claude to help me with this blog post, what limitations should I be aware of (particularly given the concerns I outlined above)? For example, I’m mindful that I’m not currently paying for this service, which could mean some sort of negative externality that would either impact on my further use of this or to raise my critical views further.
Claude:
I’m glad you’ve found our exchange valuable. When considering using Claude for your blog post, here are some important limitations to be aware of, particularly given your thoughtful concerns:
Technical limitations:
1. Knowledge cutoff: My knowledge was last updated in April 2024, so any more recent developments, research, or perspectives won’t be included.
2. No web searching: Unlike some AI systems, I can’t search the internet to find current information or specific references you might need.
3. Creative constraints: As we’ve explored, there’s still a qualitative difference between AI-generated and human-written creative content.
4. Limited context window: Very long, multi-part conversations may lose earlier context.
Ethical/societal considerations:
1. Environmental impact: Large language models do have an environmental footprint, primarily from the energy used in training and inference. While Anthropic works to minimize this impact, it’s a valid concern.
2. Labor considerations: The development of AI systems relies on human labor throughout the supply chain, including data annotation, evaluation, and infrastructure maintenance.
3. Potential overreliance: Becoming overly dependent on AI assistance for creative work could potentially affect your own creative process over time.
Usage considerations:
1. Free tier limitations: While this conversation is currently free, there may be usage limits that could affect extensive work on longer projects.
2. Data usage: Your interactions may be used to improve the model (though with privacy protections in place).
If you’re concerned about negative externalities, you might consider:
- Using Claude selectively for specific aspects of your work where it adds most value
- Maintaining a critical perspective on the outputs
- Periodically assessing whether the assistance is genuinely enhancing your writing process
Would you like more specific information about any of these aspects as you consider using Claude for your blog post?
Me:
Not at this stage, thank you, but I expect I will return. Any practical restrictions I should be aware of for support in writing my intended blog post eg an ability to enter data from other file formats?
Claude:
For using Claude to help with your blog post, here are the key practical limitations to be aware of:
1. Text-only input: Currently, you can only input plain text. Claude can’t directly access or process:
— Images or graphics
— PDFs or other document formats
— Spreadsheets or data files
— Audio or video files
2. Formatting limitations: While Claude supports markdown formatting in responses, there are limits to:
— Complex formatting needs (like precise layout control)
— Direct HTML editing
3. Length constraints: There are limits on input length and the total conversation size, though these are typically generous for blog post work.
4. No file saving/exporting: Claude can’t directly save your work or export it to other formats — you’ll need to copy text manually.
For your blog post workflow, this means:
- You’ll need to manually copy in any content from your other files or sources
- Any research materials will need to be summarized as text
- You’ll need to copy out any generated content you want to keep
If you’re planning to develop complex visuals, data analysis, or need to work with specific file formats, you might want to use Claude in conjunction with other tools specifically designed for those tasks.
Me:
Thank you — all very helpful. I note that I have one message remaining until 19:00 so a final question from me for now. If I copy the entire above exchange as a (different) blog post, a) are there any technical barriers to my copying the text out from this exchange and b) what are the copyright implications? In other words, do I have the legal right to publish this exchange as the sole author or would I co-credit Claude/Anthropic as a co-author?
Claude:
You’re welcome! Regarding your final questions:
a) Technical barriers: There are no technical barriers to copying out our conversation. You can simply select and copy the text from this interface. There’s no DRM or technical restriction preventing you from doing so.
b) Copyright implications: This is an interesting area. Here’s what you should know:
- You own the copyright to the content you’ve written in this conversation
- Anthropic maintains that outputs from Claude are not copyrightable by Anthropic (unlike some other AI companies’ approaches)
- Generally speaking, you can use the outputs from our conversation in your blog post without legal obligation to credit Claude or Anthropic
However, there are some considerations:
1. While not legally required, crediting Claude as a tool you used (similar to how you might mention using Photoshop for images) would be transparent to your readers
2. Best practices in this emerging area suggest disclosing AI use, especially for published content
3. The legal landscape around AI-generated content is still evolving
If you’d like definitive guidance on legal rights for commercial publication, consulting with a legal professional familiar with emerging copyright law around AI-generated content would be advisable.
