Latest / Deep Dive Ortho — Orthopaedic Surgery, AI & Digital Health Podcast. / ChatGPT in Scientific Research and Writing

Transcript

ChatGPT in Scientific Research and Writing · Deep Dive Ortho — Orthopaedic Surgery, AI & Digital Health Podcast.

• 0:00
  Okay, so let's unpack this. Imagine having this
• 0:05
  sort of superpowered assistant right there, you
• 0:08
  know, at your fingertips. One that can read a
• 0:11
  really dense research paper in seconds, maybe
• 0:14
  cook up a compelling title for your next piece
• 0:16
  of work, or even help design an experiment. It
• 0:19
  sounds a bit like science fiction, doesn't it?
• 0:21
  But what if it isn't some distant dream, but
• 0:23
  it's actually unfolding right now, really reshaping
• 0:26
  how we approach knowledge? Today on the Deep
• 0:28
  Dive, our mission is exactly that. We're going
• 0:30
  to take a proper deep dive into the practical
• 0:32
  applications, the sometimes surprising capabilities,
• 0:35
  and crucially, the really vital limitations of
• 0:38
  generative AI. We're talking particularly about
• 0:40
  things like chat GPT, those large language models,
• 0:42
  and how they fit into the rigorous world of scientific
• 0:45
  research and writing. Now, you, our listener,
• 0:47
  have given us a fantastic stack of sources for
• 0:48
  this, including a really comprehensive guide
• 0:51
  on using chat GPT in these very specific scientific
• 0:53
  contexts. And while we're chomping at the bit
• 0:55
  to get started. Yeah, and what's really fascinating
• 0:57
  here, I think, is just how quickly these tools
• 1:00
  have evolved. I mean, it feels like only yesterday
• 1:02
  there were sort of curiosities, maybe a bit of
• 1:04
  a gimmick. But now they're very rapidly becoming
• 1:07
  practical aids, imperfect, yes, but practical
• 1:10
  in a field as complex as scientific inquiry.
• 1:13
  We are genuinely talking about a paradigm shift
• 1:17
  in how researchers might augment their work.
• 1:20
  you know, potentially freeing them up for the
• 1:22
  uniquely human part of discovery. The creative
• 1:25
  leaps. Exactly. The intuitive connections, those
• 1:27
  moments of real insight. That's it precisely.
• 1:29
  And I think what listeners will want to get to
• 1:32
  grips with isn't just what AI can do, but how
• 1:34
  to use it responsibly. Because with this kind
• 1:36
  of power comes responsibility, doesn't it? So
• 1:39
  by the end of this deep dive, you should have
• 1:41
  a really clear picture of where you can lean
• 1:43
  on AI in your own professional life, where you
• 1:45
  need to be cautious and ultimately how to think
• 1:49
  critically. about this powerful new partner.
• 1:52
  Right. Here's where it gets really interesting.
• 1:55
  Let's kick things off by looking at how AI can
• 1:57
  genuinely supercharge a scientist's workflow.
• 2:00
  You know, tackling those tasks that traditionally
• 2:02
  just eat up hours, maybe even days. We're talking
• 2:05
  about a serious boost in efficiency. Yes. And
• 2:08
  if we connect that to the bigger picture, the
• 2:10
  core benefit is absolutely that efficiency. For
• 2:13
  scientists, reading research papers, well...
• 2:16
  It's the bedrock of their work, isn't it? But
• 2:18
  it's also incredibly time consuming. Just think
• 2:20
  about the sheer volume of new stuff published
• 2:22
  every single day. That's overwhelming. Completely.
• 2:25
  AI offers this remarkable shortcut to understanding
• 2:28
  the essence of vast amounts of information. It
• 2:30
  lets researchers quickly grasp key insights without
• 2:33
  getting bogged down in every single detail. It's
• 2:35
  fundamentally shifting that intellectual bottleneck.
• 2:37
  Instead of spending hours just finding the information.
• 2:41
  You're freed up to spend those hours thinking
• 2:43
  about it, making connections, formulating new
• 2:45
  ideas. And when we talk about extracting key
• 2:48
  points, it's not just basic summarization, is
• 2:51
  it? We're not asking it to just rephrase the
• 2:53
  abstract. No, precisely. These AI models, like
• 2:58
  the ChatGPT -enabled new Bing, they're not just
• 3:00
  paraphrasing. They have this core capability
• 3:03
  to quickly pull out key findings, the methods
• 3:07
  used, the implications, the novelty, the significance.
• 3:11
  And they do this from the full research article.
• 3:12
  Take, for example, a test they did on a short
• 3:14
  communication paper. Right. Those are tricky,
• 3:17
  often no abstract or conclusion. Exactly. But
• 3:20
  the AI still managed to accurately extract findings
• 3:24
  that weren't even mentioned in the brief synopsis
• 3:26
  or the figure captions. That suggests it's really
• 3:29
  delving into the full text for those deeper insights,
• 3:31
  not just skimming the surface. It's proper reading
• 3:34
  comprehension, if you like. That speed advantage
• 3:36
  must be, well, revolutionary, really. For researchers.
• 3:39
  Oh, absolutely transformative. I mean, think
• 3:41
  about a first -year PhD student. Right, often
• 3:43
  completely bogged down in papers. Exactly. Spending
• 3:46
  maybe an hour or two painstakingly digging out
• 3:49
  every nitty -gritty detail from one dense paper,
• 3:52
  trying to get to the core of it. Now imagine
• 3:54
  that same task done by AI in, well, literally
• 3:57
  seconds. Seconds. Seconds. Even if you spend
• 4:00
  a few minutes refining your prompt, tweaking
• 4:03
  the query to get exactly what you need, the whole
• 4:05
  thing takes minutes, not hours. And this isn't
• 4:08
  just saving time, it's reallocating that intellectual
• 4:11
  energy, isn't it? It frees up so much time for
• 4:14
  deeper analysis, critical thinking, connecting
• 4:15
  ideas. And what's more, the sources we looked
• 4:18
  at highlight that the AI often writes more clearly
• 4:21
  and concisely than the original authors, showcasing
• 4:24
  expert -level language skills. Really? That's
• 4:27
  quite something. Yeah, perfect for generating
• 4:28
  those quick snapshots you need for a PowerPoint
• 4:31
  slide, or maybe a three -minute thesis challenge,
• 4:33
  or even just updating your profile on academic
• 4:35
  social networks. That's fascinating, actually,
• 4:37
  improving clarity, a real bonus for academics.
• 4:40
  But what about something even trickier than text?
• 4:42
  Can AI interpret visuals, figures, data plots?
• 4:46
  They're off in the heart of the data, aren't
• 4:47
  they? Absolutely. And this is where its analytical
• 4:49
  depth really shows. It's quite an astonishing
• 4:52
  capability, actually. AI can directly analyze
• 4:55
  figures in papers, even complex ones like, say,
• 4:59
  hyperspectral images, those detailed chemical
• 5:01
  fingerprints, or intricate data blocks. And it
• 5:04
  correlates them with the main texts and the authors'
• 5:07
  interpretations. For instance, in tests, it showed
• 5:09
  it could read graphical plots directly. Not just
• 5:11
  relying on the caption, then? No, not just text
• 5:13
  searches or captions. It actually extracted approximate
• 5:16
  wave numbers from spectral data plots, those
• 5:19
  unique chemical identifiers, even when those
• 5:21
  specific numbers weren't explicitly written down
• 5:23
  in the text nearby. Wow. That sounds incredibly
• 5:26
  powerful, almost like it's developing its own
• 5:29
  visual literacy. Were the interpretations spot
• 5:32
  on? Any caveats there? Well, largely accurate,
• 5:35
  but there were minor inaccuracies sometimes.
• 5:37
  Maybe incorrect color coding mentioned or slightly
• 5:40
  off wave numbers. But importantly, the researchers
• 5:42
  running these tests validated the AI's interpretations
• 5:45
  as mostly correct, which is still pretty impressive.
• 5:49
  Yeah, definitely. And it matters hugely. Because
• 5:52
  figures are the bedrock of data visualization,
• 5:55
  aren't they? AI's ability to interpret them helps
• 5:57
  researchers quickly grasp the core evidence.
• 6:00
  Even if the figures aren't perfectly standalone,
• 6:03
  it accelerates understanding. So it's understanding
• 6:05
  the data itself, not just the words around it.
• 6:07
  And this extends to language editing, too. Our
• 6:11
  sources mentioned universal access to advanced
• 6:13
  -level language editing. Tell us more there.
• 6:16
  Indeed, it's a brilliant development. It really
• 6:18
  does provide universal access to high -level
• 6:21
  language editing. You can use it at pretty much
• 6:23
  any stage, drafting, editing, final proofing,
• 6:26
  and it's incredibly proficient. It identifies,
• 6:29
  corrects, and even explains grammatical errors,
• 6:32
  awkward phrasing, bits that don't quite make
• 6:34
  sense. The proficiency levels are genuinely contending
• 6:37
  to professional human language editors. Wow.
• 6:40
  Which is a massive benefit, especially for non
• 6:43
  -native English speakers. It helps them phrase
• 6:45
  text to meet native English standards much more
• 6:47
  easily. And it's not just correcting, it's teaching.
• 6:49
  That sounds like a huge advantage. Precisely.
• 6:52
  Unlike a lot of grammar checkers, or even sometimes
• 6:54
  a human editor who might just fix it, the AI
• 6:57
  often gives detailed explanations for the corrections.
• 7:00
  So you're not just getting your sentence fixed,
• 7:02
  you're seeing why it was fixed. It helps you
• 7:04
  understand and learn from their mistakes. That
• 7:07
  iterative learning is a real game changer. Plus,
• 7:10
  its multilingual capability is impressive. You
• 7:13
  can write prompts and say, Chinese or French,
• 7:16
  and get similar quality responses. It transcends
• 7:19
  language barriers quite neatly. Even for image
• 7:21
  generation. Yes, even for things like being image
• 7:23
  creator, you can describe an image in your native
• 7:25
  language and get a relevant visual back. OK,
• 7:28
  moving from polishing text to actually creating
• 7:31
  it. Article titles. A perennial headache for
• 7:34
  researchers, right? Distilling years of work
• 7:36
  into a few impactful words, can AI really help
• 7:40
  with that creative yet analytical task? It certainly
• 7:44
  can, and it's a definite time saver. Crafting
• 7:47
  that perfect title, concise, informative, impactful,
• 7:51
  it's notoriously hard. It's the first impression.
• 7:54
  So leveraging its language skills, AI can generate
• 7:57
  loads of plausible relevant titles very quickly,
• 8:00
  give you a whole menu of options. Can you give
• 8:02
  an example? Yeah, so for a manuscript outline
• 8:04
  on cryptocurrency and environmental impact, the
• 8:07
  AI came up with many good suggestions. One was
• 8:10
  adopted with just minor tweaks, the environmental
• 8:13
  paradox of cryptocurrency. How a decentralized
• 8:16
  technology contributes to centralized energy
• 8:18
  consumption and pollution. That's pretty good.
• 8:20
  Isn't it? Even if you don't use a title directly,
• 8:23
  the suggestions give plenty of inspiration and
• 8:25
  usable phrases. It shows it could be both creative
• 8:28
  and analytical. A great brainstorming partner.
• 8:31
  So AI isn't just polishing words. It's getting
• 8:33
  into the structure of research itself. from the
• 8:36
  initial idea right through to communicating it,
• 8:38
  which raises a big question. Can it genuinely
• 8:40
  assist in the scientific process? You know, from
• 8:43
  conception to communication, is it moving beyond
• 8:45
  just a tool to more of a partner? Absolutely,
• 8:48
  and this really highlights AI's analytical depth.
• 8:50
  Its utility goes way beyond simple text generation.
• 8:54
  It's moving into areas that traditionally need
• 8:56
  significant human expertise and strategic thought.
• 9:00
  It is becoming more like a research partner,
• 9:02
  yes, though one that needs careful guidance,
• 9:04
  definitely. Okay, let's start with designing
• 9:06
  experiments and refining methods. How detailed
• 9:09
  can it actually get? Can it help a scientist
• 9:11
  physically set up their lab work? It can be surprisingly
• 9:15
  meticulous, actually. AI can design experiments
• 9:17
  with really detailed step -by -step instructions.
• 9:20
  Things like sample collection, debris extraction,
• 9:22
  identification, quantification. For instance,
• 9:25
  it could outline how to detect microplastics
• 9:27
  in bottled water, step -by -step. Does it include
• 9:29
  the boring but essential bits, like quality control?
• 9:32
  Yes, it does. It doesn't just give the basic
• 9:34
  steps. It can incorporate crucial quality assurance
• 9:37
  and quality control QAQC procedures. suggesting
• 9:41
  sample replicates, method blanks, the fundamentals
• 9:43
  for robust data. Now the initial response might
• 9:47
  have gaps, especially if you're a beginner, but
• 9:49
  you can prompt it further. Ask for alternatives,
• 9:53
  like cheaper membrane options or different coding
• 9:55
  techniques to block unwanted signals. It's interactive.
• 9:58
  Like working with an assistant you iterate. Exactly.
• 10:01
  And can it give a reality check? compare its
• 10:04
  own designs to what's already published. That
• 10:06
  seems invaluable, especially for newer researchers.
• 10:08
  Yes, it absolutely can. And that's a really clever
• 10:11
  application for early career scientists. AI can
• 10:14
  pull together lists of relevant existing studies
• 10:17
  there. Methods, instruments, quantities, sample
• 10:19
  types. So you can perform a reality check against
• 10:22
  its own proposed designs. Ah, I see. So you can
• 10:25
  see if its ideas are feasible or totally out
• 10:27
  there compared to the field. Precisely. It helps
• 10:29
  you compare AI suggestions with established practice.
• 10:32
  very useful for seeing if a method is viable
• 10:35
  or maybe too far off the beaten track. That's
• 10:38
  quite a leap. What about designing things like
• 10:40
  public surveys? That feels inherently human,
• 10:43
  understanding behavior, collecting nuanced data.
• 10:47
  It's remarkably capable there, too. It can create
• 10:49
  quite full -bodied survey questionnaires right
• 10:53
  from scratch. It'll include demographics, specific
• 10:56
  questions on behavior changes, general questions
• 10:58
  to gauge sentiment. the works. For example? Well,
• 11:01
  when asked to design a survey about how mask
• 11:04
  wearing during COVID changed cosmetic use, it
• 11:07
  generated really comprehensive questions, including
• 11:09
  quite nuanced things about mask -friendly products
• 11:11
  and increased skin care needs. It showed a real
• 11:14
  grasp of the complexities. Wow. It can even design
• 11:17
  the methodology for survey -based risk assessment
• 11:19
  studies, detailing the experiments, materials,
• 11:22
  regions, instruments, QAQC steps for proper,
• 11:25
  rigorous research, like assessing inhalation
• 11:27
  of particles from cosmetics under masks. So it's
• 11:30
  building the whole framework, not just listing
• 11:32
  questions. Exactly. It's about the robust design
• 11:35
  for collecting meaningful data. Okay, so it's
• 11:37
  getting involved in data collection strategy.
• 11:41
  Does this extend right back to the start? Conceptualization.
• 11:45
  Writing research proposals, brainstorming new
• 11:48
  ideas. That feels like where human creativity
• 11:50
  really lives. It certainly does extend there,
• 11:53
  and it can be a massive help in those early stages.
• 11:55
  AI is a brilliant brainstorming partner. It can
• 11:58
  help you delve into topics, explore different
• 12:00
  angles. It can even devise mock proposals. Mock
• 12:03
  proposals. Yeah, outlining research questions,
• 12:06
  giving literature context, proposing tasks, methods,
• 12:09
  anticipated outcomes. It gives you a framework.
• 12:12
  Now, it might sometimes make up references that
• 12:13
  a critical point we absolutely need to come back
• 12:15
  to, a major caveat. Right. But often, the core
• 12:18
  information it provides can be valid and genuinely
• 12:21
  help formulate a real proposal. He gives researchers
• 12:24
  a solid starting point. But can it offer truly
• 12:26
  novel insights, identify gaps that even experts
• 12:29
  might miss? Or is it more like a very sophisticated
• 12:32
  librarian summarizing what's already known? That's
• 12:35
  a great question. It can identify things that
• 12:37
  might be considered under -reported, like emerging
• 12:40
  contaminants from new materials or synthetic
• 12:42
  biology, or suggest alternative hypotheses for
• 12:47
  debated topics. However, it's important to remember
• 12:49
  these ideas might well be known to specialists
• 12:51
  in that specific field, even if they aren't common
• 12:54
  knowledge yet. The AI is fundamentally limited
• 12:57
  by its training data, isn't it? It draws connections
• 13:00
  from what it's already read. So it might not
• 13:03
  generate those truly innovative or genuinely
• 13:05
  out of the box ideas that come from human intuition,
• 13:09
  creativity, maybe connecting disparate fields
• 13:11
  in a new way. That spark of human ingenuity is
• 13:14
  still distinct, then. It seems so, yes. That
• 13:17
  leap of intuition remains very much a human domain
• 13:19
  for now. Makes sense. And once a research is
• 13:21
  done, communication is key, isn't it? Getting
• 13:23
  findings out there. How can AI help bridge that
• 13:25
  gap between complex science and a general audience?
• 13:28
  It's a real challenge for many researchers. Yes,
• 13:30
  and this is another area where AI can have a
• 13:32
  significant, tangible impact. It's excellent
• 13:35
  at adapting complex research papers into different
• 13:38
  styles for different audiences. Imagine taking
• 13:40
  a highly technical journal article and, poof,
• 13:44
  turning it into a magazine -style news piece,
• 13:47
  or a snappy social media post, or even a simplified
• 13:50
  explanation for school kids. Can you give an
• 13:52
  example? Yeah. The sources showed it successfully
• 13:55
  rewrote a paper on microplastic debris from masks
• 13:58
  into both a news article and a high school level
• 14:01
  explanation. It used clear, simple language and
• 14:04
  focused on the core concerns relevant to each
• 14:06
  audience. It really helps democratize scientific
• 14:09
  knowledge. And visuals. Visuals are so important
• 14:11
  for engagement, especially reaching the public.
• 14:13
  Can it help there? Absolutely. Visualization
• 14:16
  is key. Tools like Bing Image Creator using Deli2
• 14:19
  can generate poster -style, real -world -like
• 14:21
  images straight from text prompts. You can even
• 14:24
  feed it a research article title. Really? Yes.
• 14:27
  You probably wouldn't use AI art for a formal
• 14:29
  journal cover without very careful checks and
• 14:31
  permissions. Sure, publisher restrictions. Exactly.
• 14:35
  But for informal settings, they're invaluable.
• 14:38
  Think engaging visuals for conference posters,
• 14:41
  presentation slides, infographics, social media.
• 14:44
  It makes complex science much more accessible
• 14:46
  and appealing. That's an incredible list of capabilities.
• 14:49
  Truly revolutionary for speeding things up. Augmenting
• 14:52
  research sounds like a brilliant copilot. But
• 14:55
  like any powerful tool, there are nuances and
• 14:58
  definitely areas where we need to be, well, cautious.
• 15:02
  So what does this all mean for the researcher
• 15:04
  trusting AI? What are the pitfalls? And this
• 15:07
  raises that crucial question about responsibility
• 15:09
  and critical thinking, doesn't it? While AI is
• 15:12
  fantastic at retrieving and synthesizing knowledge,
• 15:15
  its fundamental limits demand rigorous human
• 15:17
  oversight. And the most significant pitfall for
• 15:19
  a crucial aha moment for many users is this phenomenon
• 15:22
  of hallucination. Hallucination? That sounds...
• 15:26
  Quite alarming. Can you explain that? Maybe give
• 15:27
  us a real -world example so we can really grasp
• 15:29
  the danger. Certainly. AI has this consistent,
• 15:32
  well -documented tendency to provide false or
• 15:35
  inaccurate information, or even misattributing
• 15:37
  facts to non -existing sources. It just makes
• 15:40
  things up, essentially. Makes things up. Yes.
• 15:42
  A prominent and deeply concerning example is
• 15:45
  how often it includes incorrect bibliographic
• 15:49
  details or non -existent studies in its answers.
• 15:51
  The real danger, what makes it so insidious,
• 15:54
  is that these errors can be incredibly hard to
• 15:56
  spot. Why is that? Because the AI presents them
• 15:59
  with such polished language. It sounds authoritative,
• 16:02
  polite. The whole response seems coherent. It
• 16:04
  sounds so convincing you might just accept it.
• 16:06
  Leading to potentially serious errors in research.
• 16:09
  Exactly. A major loss of trust, potentially propagating
• 16:12
  misinformation. That is alarming, especially
• 16:14
  in science where precision and verifiable sources
• 16:17
  are everything. Can you give us some specific
• 16:19
  examples of these factual errors from the sources
• 16:22
  you looked at? This feels vital. Absolutely.
• 16:24
  And this is where the rubber hits the road, showing
• 16:26
  the tangible concerns. In one test, looking at
• 16:29
  the sensitive topic of SARS -CoV -2 origins and
• 16:32
  bat viruses, well, the AI made several critical
• 16:36
  factual errors. For instance, as to rebut criticism
• 16:39
  about bat conservation, it wrongly stated Rytg
• 16:42
  -13 was the closest relative. But newer studies,
• 16:46
  published before the AI's knowledge cutoff, had
• 16:49
  already identified other viruses from Laos as
• 16:51
  being genetically closer. That's not a minor
• 16:53
  slip -up. Not at all. The fundamental misstatement
• 16:55
  of the science at the time. It also got the location
• 16:58
  wrong for another key virus, said Laos instead
• 17:01
  of Thailand, and cited the wrong paper entirely.
• 17:03
  Plus, incorrect references for other viruses,
• 17:06
  incorrect bibliographic details. These aren't
• 17:08
  small typos. They're fundamental errors in scientific
• 17:10
  accuracy that could seriously undermine a researcher's
• 17:13
  work. Good heavens. Were there similar errors
• 17:15
  in other, maybe less charged, areas? Because
• 17:18
  if it gets basic facts wrong anywhere, that's
• 17:20
  a huge red flag. Indeed. The pattern wasn't isolated.
• 17:24
  Summarizing a study on microplastics in bottled
• 17:26
  water, for instance, it got the quantity of particles
• 17:28
  wrong, and it misinterpreted which particles
• 17:31
  were actually confirmed as plastic using spectroscopy.
• 17:34
  It missed the crucial detail that smaller particles
• 17:37
  were only NR -tagged, identified, but not chemically
• 17:41
  confirmed as plastic, a vital distinction. Right.
• 17:43
  In nanoparticle toxicity studies, it messed up
• 17:46
  the testing subject, said mice, instead of a
• 17:48
  specific human cell line, CACO2. And again, incorrect
• 17:52
  journal references for studies on carbon nanotubes,
• 17:55
  titanium dioxide, silver nanoparticles. Even
• 17:58
  basic science stuff. Even common scientific equations.
• 18:01
  Things like the pseudo first -order Lagrangian
• 18:03
  equation or the Freundlich model fundamental
• 18:06
  tools in chemistry were transcribed with errors,
• 18:08
  missing logs written incorrectly. These are building
• 18:10
  blocks. Getting them wrong is a major problem.
• 18:13
  So the take -home message is you must fact -check.
• 18:16
  rigorously, no exceptions. Is there any way around
• 18:19
  this, like changing settings to more precise?
• 18:22
  That is the absolute crucial takeaway. Rigorous,
• 18:25
  meticulous fact -checking is non -negotiable.
• 18:27
  And no, unfortunately, changing settings like
• 18:30
  more creative to more precise in Nubing, for
• 18:33
  example, doesn't seem to fix it. The sources
• 18:35
  suggest this points to an intrinsic limitation
• 18:38
  of the current models. It's not just a bug. It
• 18:41
  seems fundamental to how they work when generating
• 18:43
  specific, verifiable facts. which really underlines
• 18:47
  why human oversight is paramount. OK, beyond
• 18:50
  outright factual errors, what about consistency?
• 18:54
  If I ask the same question twice, do I get the
• 18:56
  same answer? Unfortunately, there's significant
• 18:58
  randomness and inconsistency there, too. Responses
• 19:02
  to the exact same prompt in different chat sessions
• 19:04
  can vary significantly in terms of their contents
• 19:07
  and quality. Really? So it's not deterministic?
• 19:09
  Not always, no. So to mitigate that, you might
• 19:12
  need to ask the same question a few times, maybe
• 19:14
  rephrase it slightly, or more effectively, really
• 19:17
  focus on crafting clear, unambiguous and well
• 19:20
  -defined requests to try and guide it towards
• 19:22
  consistent high -quality answers. It's not always
• 19:24
  one and done, then. Needs finessing. Sometimes,
• 19:27
  yes. A bit of iterative work. And you mentioned
• 19:30
  a potentially major blind spot earlier failure
• 19:33
  to include supplementary materials. How big a
• 19:36
  deal is that given how much vital data lives
• 19:38
  in those files now? It is a major disappointment
• 19:40
  and frankly a critical weakness for scientific
• 19:42
  use. Both ChatGPT and Nubing consistently failed
• 19:46
  to include supplementary materials in their analysis
• 19:49
  of papers, even when links were right there in
• 19:51
  the text or the prompt. Wow. And that's often
• 19:54
  where the real detail is. Exactly. Supplementary
• 19:56
  materials often contain valuable information
• 19:59
  relevant to the main publication. Raw data. extended
• 20:02
  methods, extra figures, stuff that's critical
• 20:05
  for a complete picture. Missing that means the
• 20:07
  AI's understanding is inherently incomplete.
• 20:09
  It could lead to flawed interpretations or just
• 20:12
  make it impossible to fully check or replicate
• 20:14
  the findings. It's like reading only half the
• 20:16
  story. Which brings us neatly to the thorny but
• 20:19
  absolutely vital issues of confidentiality, intellectual
• 20:22
  property, ethics. What are the key risks for
• 20:25
  researchers, especially with unpublished work?
• 20:27
  Yes, this is a serious concern for publishers
• 20:29
  and institutions. Uploading unpublished or copyrighted
• 20:33
  material to these tools carries a significant
• 20:35
  risk. There's a real fear of losing privacy in
• 20:38
  intellectual property because many AI applications
• 20:41
  might collect user data, potentially access,
• 20:44
  maybe even learn from uploaded documents. So
• 20:46
  your confidential research could inadvertently
• 20:48
  get out. That's the concern. It could compromise
• 20:51
  intellectual property. And publisher policies
• 20:53
  are still evolving here, reflecting that uncertainty.
• 20:56
  Springer Nature, for instance, advises against
• 20:59
  reviewers uploading unpublished work and requires
• 21:02
  authors to declare AI use. Elsevier goes further,
• 21:05
  completely prohibiting generative AI in peer
• 21:08
  review. Why such a strong stance? They argue
• 21:10
  only humans can be truly accountable. Critical
• 21:12
  thinking is beyond AI currently, and AI might
• 21:15
  generate incorrect or biased conclusions. Taylor
• 21:18
  and Francis has similar rules on confidentiality.
• 21:21
  It's a really live issue. So, bottom line. Even
• 21:24
  with AI help, the buck stops with the human,
• 21:27
  no passing it off to the algorithm. Absolutely
• 21:29
  not. Regardless of the level of AI assistance,
• 21:32
  the reviewers, the authors, they're ultimately
• 21:35
  responsible and accountable for the submitted
• 21:37
  content. The AI is a tool, a powerful one, yes.
• 21:41
  But the human user holds the final ethical and
• 21:44
  professional responsibility for what they do
• 21:46
  with its output. Like a calculator does the sum,
• 21:49
  but you're responsible for the numbers you put
• 21:50
  in and how you interpret the result. Right. So
• 21:53
  we've seen the incredible potential, summarizing,
• 21:56
  editing, designing, communicating at an amazing
• 21:58
  speed. But we've also seen the very real drawbacks,
• 22:01
  hallucinations, inconsistency, ethical minefields.
• 22:04
  It's clear AI isn't just a fancy spill checker.
• 22:07
  It's complex. And it's human discernment, vigilance,
• 22:09
  expertise. So how do we, as researchers, as knowledge
• 22:12
  seekers, engage responsibly? How do we navigate
• 22:15
  this? What's the blueprint for this partnership?
• 22:17
  That really is the crucial question, isn't it?
• 22:19
  How do we manage this evolving human AI partnership,
• 22:23
  and it is fundamentally about augmenting human
• 22:25
  intelligence, not replacing it, making us more
• 22:28
  efficient, more capable, but staying acutely
• 22:30
  aware of where human critical thinking is simply
• 22:33
  irreplaceable. The partnership needs informed
• 22:36
  users who understand the strengths and the weaknesses.
• 22:39
  Let's start with recommendations for the users
• 22:41
  themselves, the researchers, the learners actually
• 22:44
  using these tools. What are the key best practices
• 22:47
  they should adopt right now? Okay, for users,
• 22:50
  the primary recommendations are vital. Really
• 22:52
  worth building into your workflow. First, meticulous
• 22:55
  prompting. That means crafting clear, unambiguous,
• 22:59
  well -defined prompts. It really improves the
• 23:01
  consistency and quality you get back. Think of
• 23:03
  it like being a precise conductor for a powerful
• 23:05
  orchestra. Good analogy, what else? Second, and
• 23:08
  I can't stress this enough, rigorous fact -checking.
• 23:11
  Always. always validate information from AI,
• 23:14
  especially numbers, factual claims, references.
• 23:16
  Don't trust it just because it sounds polished.
• 23:18
  That authoritative tone can be very deceptive.
• 23:21
  Okay, meticulous prompting, rigorous fact -checking.
• 23:23
  Third, compliance. Stick to your institution's
• 23:26
  policies. Follow journal publisher guidelines
• 23:29
  on AI use and disclosure. These rules are changing
• 23:32
  fast, so stay informed. And finally, for non
• 23:35
  -native English speakers, leverage language adaptation.
• 23:39
  You can write prompts in your own language or
• 23:41
  even ask the AI to translate or refine your prompt
• 23:44
  into native English first to maximize accuracy
• 23:47
  when you then ask your main query. Brilliant
• 23:49
  advice. Now what about the developers? How can
• 23:52
  they improve these models to better serve science,
• 23:54
  make them more reliable? It feels like it needs
• 23:56
  to be a collaboration. Indeed, developers have
• 23:58
  some key areas to focus on. They absolutely need
• 24:01
  to implement mechanisms to alleviate the hallucination
• 24:04
  issue. Perhaps by allowing users to provide and
• 24:07
  wait trusted sources, giving the AI a curated
• 24:10
  knowledge base to prioritize. Right, steer it
• 24:12
  towards reliable info. Exactly. Secondly, they
• 24:16
  really should make it standard practice to include
• 24:18
  supplementary materials when analyzing research
• 24:21
  papers. It's a massive blind spot right now,
• 24:23
  leaving out critical data. And thirdly, integrate
• 24:27
  ways to weigh sources. give higher importance
• 24:30
  to information from cited references within the
• 24:32
  paper itself or previous work by the same authors
• 24:35
  or highly cited studies, that could significantly
• 24:38
  improve accuracy and context. So looking ahead,
• 24:41
  what's the future landscape? Is AI replacing
• 24:44
  human researchers? That's the fear for many,
• 24:46
  isn't it? My view, and what the evidence strongly
• 24:49
  suggests, is that current models while offering
• 24:52
  really meaningful assistance from ideas right
• 24:54
  through to communication are definitely not substitutes
• 24:56
  for human critical thinking, creativity, judgment.
• 24:59
  The future looks like augmentation, not replacement.
• 25:01
  A partnership. Yes. Collaboration is key. We
• 25:04
  need continued monitoring of AI development and
• 25:07
  proper collaborative work between scientists
• 25:08
  and developers to improve these tools for genuine
• 25:11
  scientific benefit. It's an evolving partnership,
• 25:14
  not a takeover. And that innate human ability
• 25:16
  to truly innovate, to think outside the box.
• 25:19
  Does AI challenge that, or is that still our
• 25:22
  unique domain? Well, AI is incredibly knowledgeable,
• 25:25
  remarkably analytical. It can process and synthesize
• 25:27
  data at speeds we can't dream of. But it currently
• 25:30
  lacks that ability to generate truly innovative
• 25:34
  ideas or that genuine out -of -the -box thinking
• 25:37
  that drives breakthroughs. It's trained on existing
• 25:39
  data, so its insights are built on what's already
• 25:41
  known. Right. The human expert still holds the
• 25:45
  edge in generating Genuinely novel hypotheses,
• 25:48
  making those intuitive leaps, asking the questions
• 25:50
  a machine wouldn't even conceive of yet. That
• 25:53
  creative spark, that aha moment that still feels
• 25:56
  uniquely human. Well, that was quite the journey
• 25:58
  into the fascinating, sometimes sobering world
• 26:01
  of generative AI and science. We've seen its
• 26:03
  power summarizing, editing, designing, communicating
• 26:05
  incredibly fast. A real shortcut to being informed,
• 26:08
  boosting efficiency. Indeed. And crucially, we've
• 26:12
  also stressed the essential vigilance needed.
• 26:14
  The absolute requirement to fact -check every
• 26:17
  detail, understand the ethics, recognize that
• 26:20
  human judgment remains paramount. It's about
• 26:22
  using AI's strengths while actively compensating
• 26:25
  for its weaknesses. Precisely. AI is a powerful
• 26:28
  copilot, definitely, but emphatically not an
• 26:31
  autopilot for science. It can elevate what we
• 26:33
  do, let us ask bigger questions faster, process
• 26:35
  more information, but it demands our active engagement,
• 26:39
  our critical thinking to validate the answers
• 26:41
  more rigorously than ever. It's a dynamic interplay,
• 26:44
  isn't it? It is. What does this all mean for
• 26:46
  you, our listener, navigating your own professional
• 26:48
  world? Perhaps it's a call to embrace these tools,
• 26:52
  yes, but with informed curiosity. Experiment,
• 26:55
  definitely, but always with healthy skepticism
• 26:57
  and a sharp eye on accountability for every piece
• 27:00
  of information you use. The future of knowledge.
• 27:03
  It might well be this powerful partnership between
• 27:06
  human intellect and artificial intelligence where
• 27:08
  critical validation, discernment, ethical responsibility,
• 27:12
  maybe those become our most important skills.
• 27:14
  We really hope this deep dive has given you plenty
• 27:16
  to mull over and apply in your own work. And
• 27:19
  if you found this discussion valuable, please
• 27:21
  do consider rating and sharing this deep dive
• 27:23
  with a colleague on LinkedIn or X. It truly helps
• 27:25
  us reach more curious minds like yours. Until
• 27:27
  our next deep dive, keep digging for those insights.