Latest / HumanPrint: How to Use AI Without Losing Yourself / Episode 14: Put the Torch Down

Transcript

Episode 14: Put the Torch Down · HumanPrint: How to Use AI Without Losing Yourself

• 0:00
  Foreign.
• 0:05
  Welcome to Human Print. I am Christine Whitmarsh, and if you
• 0:09
  are open to collaborating with AI, but you're not interested in disappearing
• 0:13
  into AI, you're in the right place. Each week we will look at
• 0:17
  one place where your authentic human signal is at risk for drift and
• 0:21
  one way to shore it up. Oh, and Human Print
• 0:25
  is brought to you by Abby AI Brand bank intelligence that
• 0:29
  helps authors, thought leaders, and enterprise teams train
• 0:33
  AI on their voice and point of view so written content stays
• 0:37
  unmistakably human, consistent and on brand.
• 0:41
  Learn more at abby.ink abbi.ink
• 0:47
  like you write with let's begin.
• 0:52
  Hi everyone, and welcome back to Human Print.
• 0:55
  So in the last episode we talked about freedom of creative process.
• 0:59
  Knowing how you write, understanding where tools that you choose to
• 1:03
  use fit inside that process, deciding what belongs to you, and refusing
• 1:07
  to explain or defend those decisions simply because a platform like
• 1:11
  substack or LinkedIn or some stranger on those platforms demands
• 1:15
  an account of how you used tools in your writing process.
• 1:19
  So today I want to look at the other side of that situation.
• 1:22
  What happens when the platform doesn't stop at questioning your process,
• 1:26
  but hands other people the tools to investigate it,
• 1:30
  Essentially to investigate you? Substack now allows readers
• 1:34
  to scan eligible posts, as I mentioned last episode. Notes, Comments, replies through
• 1:39
  an AI detection system called Pangram,
• 1:41
  LinkedIn has also added a reporting option labeled seems like
• 1:45
  AI slop, allowing users to flag posts they believe appear
• 1:50
  AI generated or overly influenced by generative AI.
• 1:53
  Let me say that phrase again. Seems like it's just so scientifically
• 1:58
  authoritative, right? LinkedIn is not asking whether the post contains
• 2:02
  false information, copied material, harassment,
• 2:05
  spam, or some other identifiable provable problem.
• 2:10
  It's asking users to make a process accusation based on how the writing seems
• 2:14
  to them. And once we start doing that, the writing itself
• 2:18
  stops being the primary object of evaluation. Was the idea useful?
• 2:23
  Was the information accurate? Did the writer support their argument? Did the story
• 2:27
  have a point? Was the post worth the time it took to read? Those are
• 2:31
  questions readers can answer because the work is right there in front of them.
• 2:35
  But whether a writer used AI, how they used it, where their original
• 2:39
  thinking entered the process, what they changed, what they rejected, and whether the tools
• 2:43
  supported or replaced their judgment are questions a reader generally
• 2:47
  cannot answer from the finished paragraph. That doesn't stop people
• 2:51
  from trying. Of course, given the nature of social media,
• 2:54
  spotting supposed AI writing has become its own little Internet
• 2:58
  sport. Have you noticed someone sees an EM dash, a clean, well organized
• 3:03
  structure, a word they dislike or associate with AI Even though the
• 3:06
  words all came from writers in the first place. Or a sentence that
• 3:10
  sounds too polished and suddenly they're reenacting the finale scene from A Few Good Men
• 3:14
  in the comments section. And then other people pile on.
• 3:17
  Oh, that's definitely AI. Oh, nobody talks like that. Look at the punctuation.
• 3:21
  I can always tell. No, you can't.
• 3:25
  You may be able to identify language that resembles common, unedited AI
• 3:29
  output. Yeah, I can too. I've spent years studying writing voices and working with generative
• 3:34
  AI, and there are patterns I recognize immediately. But recognizing
• 3:38
  a pattern is different from knowing the process that produced a particular
• 3:41
  piece of writing. You don't know what went into it. A human writer
• 3:45
  can sound generic. A human writer can overuse EM dashes,
• 3:49
  if that's even possible. I love EM dashes. They're awesome. A human writer
• 3:54
  can produce rigid structures, predictable transitions, and painfully polished
• 3:58
  sentences. Some people wrote that way long before Chat GPT
• 4:02
  came along, and I have the old drafts to prove it. And by the way,
• 4:05
  libraries have classic literature to prove it. AI can
• 4:09
  also participate in writing that remains highly specific to the person
• 4:13
  behind it, because that person supplied the ideas,
• 4:16
  stories, expertise, language, editorial direction, and final judgment.
• 4:19
  The human stuff. So when someone announces in the comments,
• 4:23
  they can always tell. What they can tell is that
• 4:27
  they dislike the writing. That's fine. You can dislike writing that's allowed.
• 4:31
  You don't need a technological explanation for every bad paragraph,
• 4:35
  though. It doesn't have to be a source of personal judgment. And this is where
• 4:39
  substack and LinkedIn concern me, because both companies have taken an already noisy,
• 4:45
  contentious argument and turned Suspicion into
• 4:48
  a platform feature. On Substack, readers can choose
• 4:52
  scan for AI text from a menu. Pangram then assigns an estimate
• 4:57
  of how much the text appears human written or AI assisted writers can
• 5:01
  publish a how I made this statement. They can disable detection
• 5:05
  on individual content, or they can report a detection error when
• 5:09
  detection is disabled. The reader sees this. They see that AI detection is unavailable.
• 5:14
  So what? By the way, on LinkedIn, users can report a post
• 5:18
  because it seems like there's that seems like, again, seems like AI slop.
• 5:22
  The post then disappears from that user's feed, and LinkedIn can use the
• 5:25
  report as feedback for its ranking and detection systems. LinkedIn has
• 5:30
  said AI use itself is not necessarily the problem, and that it's
• 5:34
  trying to reduce low quality inauthentic material. I understand the
• 5:38
  complaint. Really I do. There's a lot of generic AI generated content online.
• 5:43
  There are people producing enormous volumes of material that they barely read, if at
• 5:47
  all, publishing false information, filling comment sections with automated
• 5:51
  responses, and using artificial fluency to create the appearance of knowledge
• 5:55
  they do not possess. Yeah, I don't enjoy reading that either. Again, personal choice,
• 6:00
  but we need to separate the problem from the proposed solution. Because there's
• 6:05
  too much low quality content does not automatically lead to Readers
• 6:10
  should investigate one another's writing process. Those are two different
• 6:14
  things. Platforms already have reasons to address spam
• 6:18
  impersonation, coordinated automation, plagiarism, misinformation,
• 6:22
  mass produced content, you know, any kind of manipulation of the
• 6:26
  platform. Those are behavioral and editorial problems with identifiable consequences.
• 6:31
  AI use alone in someone's process does not tell
• 6:35
  us which one of those things. If anything happened,
• 6:38
  A writer could use AI during an extensive thoughtful process and publish something
• 6:43
  original, accurate and valuable. Another person could write every word
• 6:47
  manually and publish manipulative garbage. Human fingers
• 6:51
  on a keyboard are not an ethical standard unto themselves.
• 6:55
  And when a platform encourages users to focus on detecting AI
• 6:59
  rather than judging quality and accountability of the
• 7:03
  author, it creates a convenient substitute for discernment.
• 7:06
  You don't have to explain why the argument failed, where the source was weak,
• 7:10
  or what part of the writing seemed empty. You can just call it slop and
• 7:13
  move on. That may feel efficient, but it's actually intellectually lazy.
• 7:18
  Now why are the platforms doing this?
• 7:21
  I cannot tell you what was said inside private meetings at substack or
• 7:25
  LinkedIn, and neither can the people confidently declaring that they understand the whole strategy.
• 7:28
  But we can look at incentives, partnerships, product decisions,
• 7:32
  and behaviors that these features reward.
• 7:35
  The first possible motive is the simplest one, of course. Money.
• 7:39
  Substack partnered with Pangram to place AI detection directly
• 7:44
  inside its platform. Pangram gains visibility and a major real
• 7:48
  world application for its product. Substack gains a feature it
• 7:51
  can position as support for trust, transparency, and human writing. At a moment
• 7:56
  when writers and readers are worried about generic content, that doesn't mean
• 8:00
  either company secretly hates AI. Substack also provides a model context
• 8:04
  protocol connection. It's called allowing eligible publication
• 8:08
  owners to connect substack data to AI assistants like Channel,
• 8:12
  ChatGPT and Claude so they can ask questions about subscribers,
• 8:16
  revenue retention, and post performance. So they're using a tool?
• 8:20
  Apparently AI it can enter through the business entrance in the
• 8:23
  back of the Substack building. The objection arrives when it might have helped one of
• 8:27
  their users write a paragraph. And no, I'm not saying those two uses
• 8:31
  are identical. They're not. But the contrast shows why AI is bad isn't
• 8:36
  the actual operating principle the platform is perhaps Deciding
• 8:40
  which AI uses support its products and positioning, while encouraging
• 8:45
  readers to scrutinize the uses happening inside a writer's private process.
• 8:48
  So imagine that Substack is using AI
• 8:53
  as a part of its process in different ways, like writers
• 8:57
  do. The second possible motive is market share and engagement.
• 9:01
  Substack began as a newsletter platform, but it has continued
• 9:06
  adding features associated with social networks like notes, recommendations,
• 9:09
  video, live streaming, direct messaging, social discovery.
• 9:13
  And now we have the reader operated AI Scanning. Scanning gives
• 9:18
  people something else to do on the Substack platform.
• 9:21
  It invites debate, accusation, defense,
• 9:24
  commentary, and more. Scanning. So imagine that it keeps users on
• 9:28
  the platform. That's a social media theme I think we're all familiar with.
• 9:32
  It may also attract users who are less interested in writing or subscribing
• 9:36
  to writers and more interested in participating in the broader AI conflict.
• 9:40
  Now, personally, I don't think that's great because Substack was
• 9:43
  supposed to be not that. Again, I cannot prove that
• 9:48
  audience expansion was the motive behind this feature. I can just observe that
• 9:52
  suspicion creates activity and activity has value to
• 9:56
  social platforms. We've watched this happen before. Like I said,
• 9:59
  most of us have been on social media a while. Platforms learn that outrage,
• 10:03
  you know, rage baiting. It keeps people clicking and then hand everybody another
• 10:07
  button. And now they have another excuse to do that. Now, the third possible motive
• 10:12
  involves data. Substack says neither Substack nor Pangram
• 10:16
  uses publisher content to train generative AI models.
• 10:19
  Substack also provides a setting that tells outside AI tools
• 10:24
  not to use a publication's content for model training, while acknowledging
• 10:27
  that outside tools must respect that same signal for it to work.
• 10:31
  So I want to be accurate here. I have not found evidence that Substack
• 10:35
  is feeding scanned publications into a generative AI training data set
• 10:39
  that violates their statement. But generative
• 10:42
  model training is not the only valuable form of data collection.
• 10:46
  When readers scan posts, writers report errors,
• 10:50
  creators publish process disclosures, and users
• 10:54
  flag content as AI, slope those actions create labels
• 10:57
  and feedback for the platform.
• 11:00
  They show what machines classify as AI, what people
• 11:03
  perceive as AI, where the two diverge, and how users
• 11:08
  respond when a piece receives a particular designation. So they
• 11:12
  get to track human behavior. LinkedIn has explicitly
• 11:16
  described its new reporting feature as human feedback that can help improve how the
• 11:20
  platform identifies and ranks content that doesn't automatically make the
• 11:24
  feature sinister. Platforms collect feedback to improve products every day.
• 11:28
  But writers should understand that when we become the AI Gotcha
• 11:32
  police free of charge, we may also be helping a company build and
• 11:36
  refine the systems that will Judge, rank, reduce, label and distribute
• 11:40
  our own work. And then there's the broader possibility raised by all of this.
• 11:44
  Platforms may place increasing value on material they perceive as distinctively
• 11:49
  human because useful original human expression is becoming
• 11:53
  harder to separate, maybe from the mass produced synthetic AI material,
• 11:57
  as it were. Pangram's own research has focused on estimating
• 12:01
  AI use across large bodies of published work, including journalism.
• 12:05
  And its newest technical report says the
• 12:09
  system is designed to identify fully generated, edited and mixed human
• 12:13
  AI text. Those are the company's findings about its own model,
• 12:18
  rather than an independent guarantee that any individual result should be treated as a
• 12:21
  verdict. So perhaps Substack wants
• 12:25
  to protect the value of human created work. Maybe it wants better classification.
• 12:30
  Maybe it wants a stronger trust narrative for readers and advertisers. Maybe it
• 12:34
  wants engagement from the AI argument. Perhaps. Maybe several incentives
• 12:39
  are operating at the same time because, you know, corporations don't usually make major product
• 12:43
  decisions for one single reason. That's not how business usually goes.
• 12:47
  But whatever the motive, the behavioral result is sitting right in front of us.
• 12:51
  Readers are being encouraged to suspect writers.
• 12:54
  Writers are being encouraged to explain themselves and their process in advance.
• 12:59
  And people who are already eager to patrol other people's AI use
• 13:03
  have now been handed tools to do so. And this is
• 13:07
  where I want people on every side of the argument to pay attention. You may
• 13:12
  hate AI generated writing. You may want no AI involved
• 13:16
  in your own process. You may believe the technology is damaging
• 13:21
  writing, publishing, education, the arts, all of the above.
• 13:24
  You can hold every single one of these positions without scanning
• 13:28
  the written and judging the written words of strangers. Because once
• 13:33
  you accept the idea that a platform should help you investigate another writer's private process,
• 13:37
  you're also accepting that the platform might start investigating
• 13:41
  yours. And perhaps you're confident that your writing will pass. Great. But what does that
• 13:46
  mean? What does passing mean? And how can that bar maybe be changed?
• 13:49
  Pangram reports a very low false positive rate in its current
• 13:54
  technical report, and the company says its model performs well across domains
• 13:58
  and mixed human AI writing. But even when a classifier
• 14:02
  performs well across a large test set, a score attached
• 14:06
  to one person's work does not explain what happened during that person's process.
• 14:11
  And it doesn't decide whether the final work is thoughtful,
• 14:14
  responsible, or the true work of the authors. The score
• 14:18
  answers the question. The model was designed to estimate people.
• 14:22
  Then they start attaching the larger conclusions to it. Like they cheated, they lied,
• 14:26
  they're not a real writer, they outsource their thinking. They should disclose this. They should
• 14:30
  admit it it's like admit you're guilty, right? That movement
• 14:35
  from statistical estimate to moral verdict is where the personal and
• 14:39
  professional damage happens. Writers have already described Substack's
• 14:44
  rollout as a witch hunt, raising concerns about false accusations and pressure to
• 14:48
  defend legitimate uses of AI in their creative process.
• 14:52
  A witch hunt, historically speaking, does not operate through
• 14:56
  careful examination and thoughtfulness. It operates through accusation,
• 15:00
  fear, social rewards, and the belief that innocence should be easy
• 15:04
  to prove. So put the torch down. Don't scan another
• 15:08
  writer's work because you're curious whether a machine agrees with your suspicion. And don't
• 15:12
  report a post just because its style irritates you. Don't accuse
• 15:17
  someone because they used a phrase you associate with ChatGPT, or because their
• 15:21
  writing is polished or because it has an EM dash, God forbid. Or because it
• 15:24
  sounds different. That person sounds different than they did three years ago. We evolve,
• 15:28
  you know, as writers. Read the work, evaluate what's there.
• 15:31
  Check the facts, look at the sources.
• 15:35
  Ask yourself whether the writer has something to say and whether they're taking responsibility
• 15:39
  for saying it. And when the writing is generic or empty
• 15:42
  or repetitive or manipulative or boring. Yeah, you can call it generic,
• 15:47
  empty, repetitive, manipulative, or boring. Those words describe your actual
• 15:51
  critical assessment of the work. AI Slop describes
• 15:55
  a suspicion about a production method. It also gives people an excuse to skip
• 15:59
  the harder work of explaining why something failed. So here's
• 16:03
  your homework for one week. Refuse to play AI detective on any
• 16:07
  social platforms. Do not scan anyone's writing. Do not accuse a writer of
• 16:11
  using AI. Do not use sounds like AI as your explanation for
• 16:15
  why a piece did not work. Instead, name the actual issue.
• 16:18
  Was the language generic? Was the argument unsupported? Did the writer
• 16:23
  repeat a familiar idea without really adding anything to it? Was the story missing?
• 16:27
  Did the piece make a claim the writer could not substantiate?
• 16:30
  Did it sound unlike the person whose name was on it? Be specific.
• 16:34
  And while you're doing that, pay attention to what the platform encourages you to
• 16:39
  report, label, scan, or classify on its behalf again for free.
• 16:43
  You're there to read, write, think,
• 16:46
  and connect with people. You do not work for the social media detection
• 16:50
  department. And there's this week's soapbox. We have
• 16:55
  one more coming next week. I swear I'll get off it eventually. So that is
• 16:59
  it for today's episode of Human Print. Thanks for listening. Until next time,
• 17:02
  keep creating for your audience and put the torch down.
• 17:14
  And that's it for this episode of Human Print. If you want a
• 17:18
  practical framework for protecting your voice amid the AI noise.
• 17:22
  Please follow the show. And before you go, take a minute and ask yourself
• 17:25
  as the AI noise grows louder and louder, which areas
• 17:29
  of yourself do you need to protect the most?
• 17:32
  I'll see you next time.