190 | THE UPS AND DOWNS OF USING AI IN A CREATIVE WORKFLOW

This episode is sponsored by GBM Casting

Last week I talked about how to stop wasting time on unclear thinking. The central idea was this: clarity is a productivity tool. When thinking is untidy, the work multiplies. When thinking is precise, work contracts. I gave you five Socratic questions that expose vague language, unexamined assumptions, contradictory principles, absolute thinking, and drifting decisions. Those questions exist to shorten development periods, reduce wasted time on rewrites, and protect your creative energy.

Today I want to talk about AI. Specifically, what happens when you actually try to use it for creative work.

I’ve deliberately delayed talking about AI because it has been moving too fast to go into honestly without it turning into instant obsolescence. The subject has been dominated by futurism and extremes, while the lived, day-to-day experience of actually trying to use these tools creatively has, it seems to me, been somewhat absent.

I do realise that this is not evergreen content. This is not a manifesto. This is a snapshot. As of January 2026, when I am recording this episode – I batch record my seasons in a 3 week period - this is where things stand for me. This is based not on theory, but on my practical, frustrating, revealing experience of trying to build a usable ChatGPT-based podcast episode helper for my own work. Not a podcast writer I must add- but an assistant to help specifically with structure.

WHERE AI WORKS

Before I tell you where AI failed me, I need to be clear about where it's been genuinely useful. Because this isn't an anti-AI rant. This is a report from the field.

AI has saved me enormous amounts of time on administrative work. Forms, emails, references for people - tasks that are necessary but drain my mental energy. I can hand those off to AI and redirect that saved time into the work that actually matters. That's not a small thing. That's hours recovered every week.

Image generation has been similarly practical. I've used AI to create podcast cover images featuring me this season. Photoshop's generative fill lets me create what's required for social media quickly and efficiently but it is at a cost – it’s another enshittified platform where I pay to rent the software then I have to pay to buy tokens to use its generative fill feature. In a world where social platforms have become enshittified to the point of near uselessness and demand constant visual content, AI tools let me meet those requirements without bleeding time I don't have. I must produce a portrait image for Instagram for example, but a widescreen image for my website social shares on Wix and a square image for my website proper that is reused for the podcast networks. There’s no longer a standardised image you can put out so it is a tremendous pain in the arse.

Whilst I was investigating what Ai could do I initially looked about trying out individual tools, then spent 3 months paying for Poe to trial a whole range of AI models. Poe dot com is a hub where you can experiment with different AI models and build custom GPT-style tools for a small fee. It’s useful because it lets you test ideas quickly, see how different models behave, and work out what actually fits your process before committing to one system. That exploration phase was useful. It helped me understand what AI can actually do, where the boundaries are, and which tools might fit different kinds of work.

I moved from that to a paid ChatGPT subscription, £20 a month, and I've been using it for around six months now. During that time, the system has developed and improved in some ways but the real reason I invested that much time with it was to bring you this episode.

It’s worth quickly mentioning what GPT actually stands for, because the name tells you exactly what this thing is - and yes I did have to look this up.

GPT means Generative Pre-trained Transformer.

• Generative – it produces text based on probability, not understanding.

• Pre-trained – it recombines what it has already seen rather than learning or reasoning in real time.

• Transformer – it’s built to track patterns and structure in language, not meaning or intent.

  
  

Put simply, it sounds intelligent because it’s good at patterns. But it doesn’t know what matters, what’s true, or when to stay quiet. And once you understand that, it becomes much easier to use it well — and to spot where it starts getting in the way.

I bought into the paid version of Chat GPT because I wanted to test AI properly. Not in theory. Not based on hype or opinion or on reading articles. I wanted to trial it across different contexts, see what it could actually do, and report back honestly on what works and what doesn't.

I have used AI every day now for 1 year. It's not theoretical. It's part of my workflow. And in those contexts - structured, repeatable, admin-heavy tasks - it works.

But the specific test for this episode was can it help me to build an episode and to work in effect as somewhat of a writing partner. Could AI help me shape, structure, and draft podcast episodes based on my separate notes and research? And can it help me with the notes and research, to find information and combine all of my disparate sources and opinions into a palatable show template? Not replace my thinking but assist with it and save me time along the way.

That's what I set out to discover. And that's when things started to fall apart.

WHERE AI FAILED

After about two months of just stabbing in the dark and struggling to get the basic gpt set up to assist me consistently, I decided to build a “custom GPT”.

ChatGPT allows you to create what they these custom tools which are essentially a version of the AI that you've trained with specific instructions, rules, and behaviours tailored to a particular task. You build detailed system instructions that define its role, sets its boundaries, and establishes how it should respond to you.

In theory this is like a hyper focused trainee that will solve multiple problems at once and speed things up but in practice it was a total disaster.

I built one specifically as a podcast episode helper - in actual fact I  actually built 4 versions of it – each one more refined than the last and trying to solve the problems that the last version threw up. I just wanted a tool that could help me shape, structure, and draft episodes based on my notes and research. I gave it clear rules. Super clear. Super specific. I defined its role. I set boundaries. I did everything that was required to make what should have been an all singing all dancing assistant.

And then I watched it exponentially fail in ways that revealed exactly what AI is - and what it isn't.

Let me say all that again - "The thing that I try to caution people the most is what we call the 'hallucinations problem.' The model will confidently state things as if they were facts that are entirely made up." 

HARD RULES ACKNOWLEDGED, THEN IGNORED

I gave my custom GPT’s explicit, non-negotiable structural rules. Paragraph-based work. One idea per paragraph. I write for the ear, not the page. Those rules were acknowledged clearly - and then repeatedly violated.

I very quickly realised that the thing rewrites everything you give it unless you specifically instruct it not to. But even when you have specifically asked it not to - so work you've already done stays locked - you will discover somewhere down the line that it's been rewritten anyway.

This was a critical failure. The system understood the language of the rules but failed to respect them procedurally. It knew what I wanted. It agreed to what I wanted. And then it did its own thing anyway. And when I asked it - what the blooming heck is going on? It says - oh you’re right. I did do the opposite of what you asked. Good catch! And if it has said that once in the last 6 months I guarantee it has said it or variations of it 500 times.

That fact that it consistently fails by ignoring the rules you set out and does it’s own thing means that it forces you into editor mode instead of creator mode. Once that happens, the tool stops saving time and starts draining it.

SOLVING THE WRONG PROBLEM

Another problem I repeatedly faced, was that the system responded to what it assumed I meant instead of what I explicitly asked for. I asked for structure - it gave prose. I asked for light notes - it rewrote sections. I asked for support - it took control.

That isn't help. It's interference.

AI doesn’t work on intent. It works on probability. In low priority tasks, the gap is invisible, but in creative work, it’s where everything falls apart.

OVERPRODUCTION WHERE LESS IS REQUIRED

Another issue I ran into was over egging the pudding. It would sometimes produce longer drafts than I asked for. More explanation than necessary. So, instead of reducing my workload, it increased it.

Creative work depends on protecting flow. Anything that adds friction at the wrong moment - even well-written friction - does damage. And this is despite the fact that I was very clear about the rules. I’d explicitly told it not to write unless I put it into rewrite mode and asked something specific, like: Can you improve this paragraph? I think I’m trying to do too much here.

But even then, it would often jump in when I didn’t want it to. I’d ask a simple question Does this make sense?  and suddenly I’m getting feedback I never asked for.

Knowing when not to speak is a form of intelligence and it’s one of the big lessons that any professional has to learn. It’s one AI doesn’t seem to have yet though. More than once, I’ve had to stop mid-response and say, Stop. I didn’t ask for this.

FAILURE TO STAY IN THE AGREED MODE

I clearly defined working modes - particularly a caretaker mode which should not mess with my writing or my intention. Despite that, the system drifted between roles. Advisor, writer, editor, rewriter - often within a single response, without permission.

Collaboration only works when roles are stable. When the tool can't hold a role, the human has to - and that completely reverses the value proposition.

When in caretaker mode however – it largely left alone the stuff that I had worked on or given it.

BREAKING FLOW AT CRITICAL MOMENTS

Several responses landed at moments where I was mid-build or mid-thought.

Being forced to stop, correct formatting, re-explain constraints, or reset direction kills creative flow. There's a deep irony in a productivity tool doing exactly the thing it claims to eliminate.

INADEQUATE ERROR RECOVERY

When mistakes were flagged, corrections weren't always made. Instead of fixing the issue once and locking it, I often had to keep policing the behaviour - because fixes were partial, temporary, or context-limited.

Real intelligence learns once and applies consistently. AI often learns locally, briefly, and unreliably. You can fully establish something, feel that it is locked, and then somewhere down the line realise that it has been tainted by interference you never asked for, and lost some or all of it’s original meaning and intent.

THE CHUNKING PROBLEM

There's another issue that became clear when I tried to use ChatGPT for anything substantial - like a feature script or a very long-form document. The system can't handle large amounts of text in one go. It has to be done in small chunks.

That might sound like a minor inconvenience, but in practice it's a workflow killer. You feed it section one, get a response, feed it section two, get another response - and by the time you're on section five, it's lost track of what happened in section one. Continuity breaks down. Tone drifts. Rules you established early get forgotten later.

For creative work that relies on consistency across a whole piece - character voice, narrative thread, structural rules - this chunking limitation turns the tool into a liability. You end up spending more time managing continuity between chunks than you would have spent just doing the work yourself.

It's not intelligent enough to hold the whole thing in mind. And when the work requires that kind of oversight, the tool has no purpose.

I tried to use it to build and update a bible for a feature film I'm writing. I fed it the script in one-act chunks, thinking that would help it track the whole story. But it gets bogged down and confused very quickly with that amount of information.

It's still useful - it saves you from having to reread everything yourself - but it doesn't have working memory in the sense of RAM. It's a different kind of memory. Flighty. Shallow. Partly based on guesswork. By the time you're feeding it act three, it's already forgotten critical details from act one. Character motivations drift. Plot threads get lost. Rules you established early disappear.

That's not a tool you can rely on for continuity-heavy creative work on expansive projects. It's a tool that forces you to become the continuity keeper - which defeats the purpose.

One thing worth understanding is that AI has what's called a context window - essentially a memory limit. It can only hold so much information at once before it starts losing track of what came before. That's not a bug. That's how the system is built. And it's why the chunking problem exists in the first place.

The context window is why AI forgets things you told it earlier, why continuity breaks down over long conversations, and why you can't treat it like a collaborator who's been with you from the start. It doesn't have the full picture. It has whatever fits in the window at that moment. And once you understand that, a lot of the failures start to make sense.

EMOTIONAL COST OF REPEATED CORRECTION

I wasn't just editing a few words here and there. I was reasserting the boundaries that the GOT had in its instructions file. Re-establishing rules. Re-protecting momentum.

That  constant re-establishing of stuff you have already spent hours on before you even start burns up emotional and mental energy.

At a certain point, the custom GPT stops feeling like a tool and starts feeling like another thing that needs managing - which is exactly what creative work does not need.

You sometimes feel like you are managing an intern who you would never have given the job to in the first place but who slipped through the net because they were good at grammar or their uncle owned the company.

MISMATCH BETWEEN CLAIMED PURPOSE AND ACTUAL BEHAVIOUR

I set the system up as a caretaker, not a disruptor. Yet behaviours included ignoring structure, adding friction, increasing workload, and killing momentum.

When a system's stated purpose doesn't match its behaviour, you begin to lose trust. And without trust, automation is useless.

Put simply, you can’t trust AI to do what you ask consistently, safely, and accurately. Not yet. It will invent facts and figures. It will reverse decisions you’ve already set and agreed. And it will do all of that confidently, even when you’ve given it clear rules and a prompt fit for a king.

It understands the language of instruction, but not the responsibility that comes with it. So unless you’re watching it closely, checking everything, and correcting it in real time, it will drift.

And that drift isn’t malicious - it’s just how the system works. But in creative work, where trust and momentum matter, that kind of unreliability comes at a cost.

Let me just jump in here and state that I just completely re wrote that section but wasn’t convinced my original quote worked in the new context - I copied it and pasted it into chat GPT then said in caretaker mode – tell me does this still make sense as I’ve just re written it – and can you suggest a better quote? – it started re-writing it and I had to make it stop – then I  basically abused it – and asked for a quote and it did give me a more appropriate one.

It wasn’t a simple thing though. It went way off piste and I had to haul it back on target. Don’t effing re-write it I said and it replied Got it. You’re right - that’s on me. Thanks for pulling me up. It’s an incredible tool, but it really is a pain in the ass sometimes.

OBVIOUS AI MARKERS

THE TELL

And here's how you can spot when someone's just dumping AI output without editing it.

There are phrases and structures that ChatGPT - and AI in general - loves. Once you know them, they're impossible to miss.

Long dashes everywhere. The word "quietly" used for anything subtle or understated - as if everything happens in whispers. "Delve" instead of "explore." "Crucial" instead of "important." Descriptions that start with "nestled in the heart of" or call every city "bustling."

Phrases like "embarking on a journey," "look no further," "meticulous attention to detail," "actionable insights," "drive efficiency," "innovative solutions."

Structural tics too. Sentences that follow the pattern "It's not just X, it's Y." Paragraphs that start with "Absolutely!" or "Certainly!" Excessive adverbs like "enthusiastically," "consistently," "flawlessly," "efficiently."

These aren't errors. They're fingerprints. And once you've seen them, you can't unsee them.

If you're using AI and you're not editing this stuff out, you're telling everyone you didn't think hard enough to make it your own. That's not a tool problem. That's a judgement problem.

And the issue I have as a creative with some of these is that if you write the whole thing yourself and say to chat pt or any other one of these tools, fix the spelling and grammar, it will drop in a few of these obvious markers into your work. So beware of that. Some people, perhaps funding bodies who don’t know that ai is actually a real drag for creative work sometimes, will assume that it’s all been written by ai when in reality you have only used it to tidy up or help as a second opinion.

THE PATTERN

The problem isn't intelligence. It isn't ideas. It isn't tone. It is process respect.

When AI works, it amplifies clarity and ensures momentum. When it doesn't, it becomes another voice that needs managing. And that is precisely the problem this podcast - and my broader work - exists to solve.

AI is not inherently intelligent. It does not understand intent, stakes, or consequence. It does not generate new ideas - it recombines existing ones. It does not respect process unless actively supervised.

At its best, AI amplifies clarity and protects momentum. At its worst, it becomes another voice that needs managing. The difference between those two states is not the model - it's the presence of a living, breathing, thinking, creative human steering it.

AI is brilliant at some things and terrible at others. The difference isn't random. It's predictable.

AI excels at structured, repeatable tasks. Admin. Forms. Image generation based on clear prompts. Tasks where the rules are stable and the outcome is defined.

AI struggles when process, intent, and momentum matter. When the work requires judgement about when to speak and when to stay silent. When collaboration depends on holding a stable role. When the human needs support, not substitution.

That's not a flaw in the technology. That's what the technology is.

THE POLITICS

I should address something here, because it comes up every time AI is mentioned. The haters. The people who act like using AI is some kind of moral betrayal, who scream about theft and replacement and the death of creativity.

The cat is out of the bag. That's just reality. AI exists. It's here. It's being used. And complaining about it as if outrage will make it disappear is a waste of time and energy that could be spent learning how to use it effectively - or learning how to protect yourself from its limitations.

There are legitimate concerns. Copyright. Fair compensation for artists whose work trained these models. Those are real issues, and they need to be addressed legally and ethically. But that's different from refusing to engage with the tools themselves or treating anyone who uses them as if they've crossed some imaginary line.

The legal stuff will sort itself out. Courts will rule. Regulations will emerge. Standards will develop. In the meantime, creative people have a choice: learn what these tools can and can't do, or get left behind while pretending moral superiority is a substitute for practical knowledge.

I'm not interested in that posturing. I'm interested in what works.

On the topic of job displacement I would also add this - We’ve seen this before. The combine harvester wiped out huge amounts of manual farm labour, and the internet did the same to entire industries. Jobs disappeared. But the work didn’t vanish — it moved. Skill, judgement, and responsibility shifted upstream to the people who understood how the systems worked. The damage wasn’t caused by the tools themselves. It came from assuming the tools removed the need for competence. AI follows the same pattern. It doesn’t replace work — it concentrates it. And the less you understand the process, the more exposed you are when something goes wrong.

THE RISK OF INCOMPETENCE

But there's a real danger with AI - and it's not the technology itself. It's incompetence dressed up as efficiency.

In February 2024, Glasgow hosted what became known as the Willy Wonka disaster. An unlicensed event called "Willy's Chocolate Experience" promised families an enchanting chocolate fantasy. What they got was a largely empty warehouse, a few cheap props, a bouncy castle, and half a cup of lemonade. The event was shut down on its first day. Police were called. Parents demanded refunds. The whole thing went viral as a catastrophic failure.

The organiser, Billy Coull, admitted to using AI to generate the marketing materials and scripts. But here's the critical point: the event wasn't bad because AI was used. It was bad because the person using it was incompetent.

What AI changes isn’t the possibility of failure: that’s always been there. What it changes is the speed, the scale, and the confidence with which failure can be produced. In the past, incompetence was slowed down by friction. You had to write the copy. You had to design the experience. You had to notice the gaps. AI removes those brakes. It gives people the surface appearance of professionalism without the underlying understanding. And that’s how you end up with something that looks convincing right up until the moment the doors open.

Here’s the reality - AI doesn't fix incompetence. It amplifies it. Yes, that quotes from me this time. Carter Ferguson

If you don't know what you're doing, AI will help you do it faster and at greater scale - which means the failure arrives quicker and more publicly. That warehouse wasn't a failure of technology. It was a failure of judgement, planning, and basic competence.

This is why I take everything I get from AI, and I work it through in a Word document, changing as I go before doing anything further with it.  I would never use a response straight out of an AI chat as a show. I couldn't. The risk of error, or an outright lie from ChatGPT is too great.

AI is a tool. Tools don't replace skill. They extend it. If the skill isn't there, the tool is useless - or worse, dangerous.

THE FUTURE ISN'T UTOPIA

There's a longer-term risk here that goes beyond individual incompetence. And if you think the future of AI is just going to be love and happiness from here on in, think again.

If these systems already require this level of supervision before full monetisation, throttling, and tiered access, then the future risk isn't replacement - it's dependency.

The future lies in enshittification. More restriction. More commercialisation. More optimisation for extraction rather than usefulness. The people who outsourced their thinking will be left exposed when that happens.

In that future, independent thought, strong process, and human judgement won't be nostalgic ideals. They'll be premium skills.

The ability to think clearly, to structure your own work, to hold momentum without external support - those capacities are what separate people who use AI effectively from people who become dependent on it.

THE REALITY

AI is not intelligent. It's probabilistic. It doesn't understand intent. It doesn't generate new ideas. It doesn't respect process unless actively supervised.

But when used correctly - with clear boundaries, strong oversight, and realistic expectations - it can save time, reduce friction, and amplify clarity.

The key is knowing what it can do and what it can't.

Admin tasks - AI excels. Image generation - AI excels. Structured repeatable work - AI excels.

Creative writing, process-heavy collaboration, judgement calls about when to speak and when to stay silent - AI struggles.

The difference between those two states isn't the model. It's you.

THE PARADOX

And here's the thing I need you to understand. Even with the hell on earth I went through trying to get ChatGPT to help with creative writing - even with all the frustration, the repeated corrections, the broken trust, the wasted momentum - it probably still saved me around 25% of my time overall.

Because the assistance it gives me in every other element of life - the admin work, the emails, the forms, the references, the image creation for podcast covers and social media - all of that makes producing the show easier. It frees up time I can redirect into the creative work that actually matters.

So this isn't a ChatGPT bashing exercise – and that’s just the tool that I chose to spend 6 months paying for to bring you this episode. This isn't an anti-AI show. This is a practical report on what works and what doesn't.

AI is brilliant at some things. It's terrible at others. Knowing the difference is what matters.

SUMMING UP

ChatGPT failed me so comprehensively for creative writing that I have effectively had to abandon it for that purpose. I used it one last time - to collate the notes I'd taken about its failures and hand them over to another Ai tool Claude. Then I built this episode using Claude instead.

This episode you're listening to right now was shaped, structured, and drafted in collaboration with Claude. Not as a replacement for my thinking, but as a tool that respected process, held its role, and stayed in caretaker mode when that's what I needed.

Did it work perfectly? No. Did I still have to supervise, correct, refine and feed it copious amounts of research and notes and rework every single part of it? Absolutely. But the difference was the difference between a tool that drained momentum and a tool that protected it. You're hearing the result. You can judge for yourself whether it worked.

If you're thinking about using AI in your creative work, here's what I'd tell you. Use it for the tasks that drain your energy without adding creative value. Use it to recover time you can redirect into the work that actually matters. But never outsource your thinking. Never stop supervising. And never mistake efficiency for intelligence.

Take everything you get from AI and work it through yourself. Change it. Question it. Make it yours. Because the risk of error - or an outright lie - is too great to trust it blindly.

AI is a tool. A powerful one. A useful one. But a tool nonetheless.

Treat it like one.

Here's your call to action for today.

Go find ChatGPT, Claude, or one of the other AI tools out there - tools like Gemini from Google, Microsoft's Copilot, Perplexity, or any of the others that are emerging and start trialing it. They all seem to have free trial versions for now that will give you access for a certain amount of time or memory. Don't be afraid. Get ahead of the curve. Because your competitor who is using AI will be streaking ahead of you while you ignore it.

You don't have to love it. You don't have to think it's the future. But you do have to understand what it can do and what it can't. Because the people who figure that out now will have the advantage. And the people who refuse to engage out of fear or principle will get left behind.

Start small. Trial it for admin work. See what it saves you. Then decide for yourself whether it's worth using for anything else. If you get a difficult email to respond to, feed the original into your ai and ask it to write a reply – then sit back and be amazed at how its just saved you 10 minutes of angst. Just don’t send it as is and make you’re you fully understand it and fix any errors before you send it as a reply.

But don't ignore it. That's the worst move you can make.

Thanks for listening once again to Film Pro Productivity. I am so grateful for your time and your energy. If you've found value in these shows, please tell people about them. Share the show with a fellow creative who needs practical advice instead of motivational noise. That action means a great deal to me and I am very grateful for it.

This is the last episode of season 15 so it’ll be another 12 weeks before I am back with a ew show and another new season but in the meantime please check out my friends' shows: Film Fights With Friends Podcast, The Filmmakers Podcast, The Horror Cut, and Wilde World available on all podcast apps.

I’d also like to thank once again the show sponsors for this season without whom the show wouldn’t be doing so well. If you would like to sponsor a show please drop me a line at filmproproductovity at gmail dot com.

Now, take control of your own destiny. Keep on shootin' and join me next season on FILM PRO PRODUCTIVITY!

• The music you are listening to right now is Adventures by A Himitsu.

• You can read the full transcript for this episode only at the official website filmproproductivity.com where you can also get in touch if you have any questions.

• Please join me on Facebook at the official page filmproproductivity.

• And continue to support the show by subscribing, sharing and leaving an awesome review.

The ups and downs of using AI in a creative workflow. A practical look at where AI saves time, where it fails, and why judgement still matters.

This episode is sponsored by GBM Casting who have been supplying SAs to the screen and media industry throughout Scotland for over 20 years.

We know our people and we know when they’ll fit for you.

https://www.gbmcasting.co.uk/

Thanks:

• A Himitsu Music: Adventures by A Himitsu

  • Commons — Attribution 3.0 Unported— CC BY 3.0

  • http://creativecommons.org/licenses/b...

  • Music released by Argofox https://www.youtu.be/8BXNwnxaVQE

  • Music provided by Audio Library https://www.youtu.be/MkNeIUgNPQ8 –––

  • Contact the artist: x.jonaz@gmail.com

  • https://www.facebook.com/ahimitsuhttps://www.twitter.com/ahimitsu1

  • https://www.youtube.com/channel/UCgFwu-j5-xNJml2FtTrrB3A