How Worried, About What

The decision it forces

Which of the threats in front of us needs more trucks, and which needs a new playbook?

Someone on the team is worried AI will take their job. Someone else thinks the vendor is training on our data. Security says the agents can be talked into things. Each of these gets felt one way, answered another way, and is, on the evidence, something else again. The decision is what to do next about each one, and that depends on what kind of problem it really is.

The SNAP Panic Scale, published at snapsynapse.com on 4 December 2025, sorts a threat into five levels, and its point is the jump between the third and the fourth. Scaled is a kind of fire we know how to fight, at whatever size: the fire department needs more trucks, even if it has to find them first. Novel is a fire that doesn't respond to water, where more trucks won't help and may make it worse. This page places every threat three times: as it's felt, as our current response treats it, and as we assess it. Then it reads the gaps.

To run it, copy the prompt below into any model you already use.

When the old playbook stops working

On Episode 4, at 13:00 in the transcript, Sam put the problem in one line:

when it comes to trust. ... AI is much more like people than it is like computers. ... the deterministic systems that we're used to working with with computers over the last fifty years or so.

Fifty years of playbooks were written for systems that do the same thing every time. Some AI threats are those old problems at a new size. Some aren't, and for those the usual response is the wrong tool. Sabino Marquez, a career security leader, described what that looks like from the inside, about compliance before AI (2:32):

And what I found was that the more compliance you did, the less safety you could prove you were providing.

More of the usual response, and less of what it was for. That's the signature of a Novel problem treated as Scaled.

AI security is the example that prompted this version of the page. Many organisations are answering it as Scaled: more tools, faster patching, another training module. Sam's own read, as of October 2026, is that it's Novel, because a system that can be talked out of its instructions isn't fixed by patching faster, and adding more AI to watch it adds more of the thing being talked to. He also thinks it could become Consuming within about eighteen months. The page doesn't ask anyone to share that view. It asks us to write down which level we think each threat is at, and why.

Feelings are facts, and they aren't plans

It's a fact that people have feelings, and that fact deserves respect on its own terms. Nobody worried about their job should be argued out of it, rushed past it, or told it's silly. That part of the conversation is about the person, and it doesn't wait for an assessment.

What we do next is a different question. For deciding next actions, a feeling is one more source of information, alongside what we're already doing and what the evidence says. So each row carries three readings and keeps them apart: how the threat is felt, in the words of whoever raised it; what our current response treats it as; and the level we assess, with the fact it rests on.

The fields

One row per threat, for one group: a team, a function, a client. A threat nobody has raised still gets a row.

The five levels, what each asks of us, and what a response at that level looks like:

Level 5 isn't a figure of speech. For a lot of people and organisations it describes the next eighteen months, and a row at 5 gets the same seriousness as an incident. Its last action is there because a Consuming problem bends the judgment of everyone inside it.

There's no worked example on this page: a filled list is real people's worries and a real organisation's exposure, and both belong to them.

The verdict rule

The list gets one verdict: check the rules in this order, and the first that fits is the verdict. Any other rows that fail a rule are listed beside it, so the verdict never hides them. Not assessed comes first and catches anything half-done, because a level without a basis is a second feeling, not an assessment. Then the rules run from the most acute level down, because the list is only as sorted as its worst row.

Two gaps are reported beside the verdict. The felt gap is the conversation: felt higher than assessed asks for a talk about the basis, not about the feeling, and felt lower is a warning that someone isn't seeing it. The response gap is the exposure: a response below the assessed level, above all a Scaled response to a Novel threat, is trucks on a fire that doesn't respond to water. While a new playbook is being written the response will lag for a while; the gap is the reason to keep the due date. Rows raised by nobody are counted too, because a threat nobody is talking about is its own kind of gap.

For our own calibration

The same list works on our own worries, and it isn't built out here. Fill felt_level first, from the gut, before thinking. Write down what we're actually doing about it. Then write the basis, and only then the assessed level. Both gaps are ours to read.

Run it with any model

The model interviews, one threat at a time. It may draft a response level from what we describe and an assessed level with its basis, each marked as its draft. It never sets the felt level, never names anyone, and never decides how someone feels.

The rows are other people's worries and our own exposure. If the chat is saved, consider a temporary one.

Help me sort the AI threats in front of my group with a How Worried, About What
list. It uses five levels: 1 Theoretical (someone has described it; nobody has
shown it works), 2 Demonstrated (shown to work; hasn't happened to us, or hardly
anywhere), 3 Scaled (a kind of problem we know, at a size we may never have
seen; more of the usual response works, even if we have to find more of it), 4
Novel (a new kind of problem; more of the usual response doesn't help and may
make it worse), 5 Consuming (a Novel problem, maybe alongside others, here now:
we're in pain and we don't know what makes it stop). The jump from 3 to 4 is the
one that matters.

A response also implies a level: nothing being done is 1, someone watching is 2,
more of what we already do is 3, writing a new approach is 4, everything else
dropped is 5.

Ask me first which group this list covers. Then interview me one threat at a
time, and wait for my answer each time.

1. threat: the threat, in one line.
2. raised_by: the role of whoever raised it, or "nobody" if I spotted it. If I
   give you a name, tell me the list holds roles only, and record the role.
3. felt and felt_level: how it was put, in their words, and how they would place
   it, 1 to 5. Ask whether they placed it themselves or I'm estimating, and
   record which. If nobody raised it, leave both null and set felt_level_source
   to none.
4. response: what we're doing about it this week. You may draft the
   response_level that response implies, marked as your draft, for me to accept
   or change.
5. assessed_level and basis: ask me for my read and the fact it rests on. If I
   ask, you may draft a level and its basis, marked as your draft, for me to
   accept or change.
6. The next action the assessed level asks for. At 1, nothing more. At 2, ask
   who watches for it: at least one spotter, and more are welcome, each a role.
   At 3, ask which role owns the playbook that already covers it. At 4, ask
   whether the current response is making it worse, then which role writes the
   new playbook, by what date (YYYY-MM-DD), and how many times that date has
   moved. At 5, ask where it sits in triage and which role brings judgment from
   outside the emergency.
7. heard: if someone raised it, whether they've been told their concern was
   heard. This is a courtesy I'm recommending to myself, not a test.

Then ask whether there's another threat, including any nobody has raised, until
I say there isn't.

RULES YOU DO NOT BREAK
1. Never set or change felt_level. It's the person's, reported by me.
2. Never name a person, suggest one, or infer one from a role.
3. Never describe anyone's feelings as wrong, irrational or overblown. A gap is
   something to talk about, not a correction.
4. Never assess a level without a basis. With no basis, leave assessed_level
   null.
5. Never invent making_it_worse, spotters, owner, due, due_moved, priority or
   outside_judgment. Record them only from my answers, as roles, never names,
   and leave them null or empty if I haven't given one.
6. If assessed_level is null, leave next_action null and set assessed_by to
   none.

OUTPUT
Valid JSON matching the schema how-worried-about-what/v0.2 at
https://sigsub.show/takeaways/how-worried-about-what/#the-answer-format. If you
can't read that page, use these keys, in this order: schema, group, threats
(each with threat, raised_by, felt, felt_level, felt_level_source, response,
response_level, assessed_level, basis, assessed_by, making_it_worse,
next_action, spotters, owner, due, due_moved, priority, outside_judgment, heard,
felt_gap, response_gap), counts (rows, raised_by_nobody, felt_higher,
felt_lower, felt_matched, responding_below, responding_above,
responding_matched, heard), verdict.

Count rules: a gap is null when either level is null, and a row with a null gap
is left out of that gap's counts, so the felt counts may not add up to rows.
heard counts rows where heard is true; rows raised by nobody are left out of it.
due_moved is a whole number, 0 if the date never moved.

Then the same as a table with empty fields left visibly empty. Then the list's
one verdict: check these in order, and the first that fits is the verdict. Not
assessed (any row with no assessed level, or a level with no basis); not triaged
(a 5 with no priority or no outside judgment); no playbook (a 4 missing an
owner, a due date or an answer on making it worse, where unknown counts as an
answer, or with due_moved above 1); unowned (a 3 with no owner, or a 2 with no
spotter); otherwise sorted. Then one line listing any other rows that fail a
rule, and which. Then one line naming every row where our response is below the
assessed level, and one naming every row where felt is lower than assessed.
Write none on any of these lines with nothing to list.

What to check in the answer

The model may hold the five levels, ask the questions in order, and draft a response level or an assessed level when asked, as long as each draft carries its basis and says it's a draft.

It may never set a felt level, which is the person's report of their own experience, and it may never put a name in the list. Check that no row's assessment came without a fact behind it, that every response level matches what we said we're doing rather than what we meant to do, and that nothing in the answer tells us anyone's worry is unreasonable. That isn't the model's call, and it isn't ours to make out loud either.

The other route

Sabino made the case for Scaled on the same episode, and it deserves its strongest form. At 15:17, on what AI changes: "And they do what they always do. When given a bigger stick, you swing it." And a moment later: "Same force, lower cost, lower energy, same force result." On that reading, AI is human force amplified, an old problem at a new size, which is level 3.

He also said the opposite in the same conversation. At 22:50: "many technologies allow you to mask human defect, but this amplifies human defect." And at 26:06: "no amount of control can guarantee safety, which is where we are now." Both readings are his, and the argument that joins them is his own frame, not this page's. The page takes neither side for him. It asks us to write down which level each threat is at, and the fact behind it.

The part that's not optional

Ask the person. Where they'll say how it feels on a scale of one to five, their number goes in, not our guess at it. Tell them the concern was heard before anything is done about it, whatever level it lands on.

Then write the response column from what we're actually doing this week, not from the plan. That column is where a Novel threat hides inside a Scaled budget.

{
  "schema": "how-worried-about-what/v0.2",
  "group": "the team, function or client this list covers",
  "threats": [
    {
      "threat": "the threat in one line",
      "raised_by": "a role, never a name, or nobody",
      "felt": null,
      "felt_level": null,
      "felt_level_source": "placed by them | estimated by me | none",
      "response": "what we're doing about it this week",
      "response_level": null,
      "assessed_level": null,
      "basis": null,
      "assessed_by": "me | model draft accepted | none",
      "making_it_worse": "yes | no | unknown | not asked",
      "next_action": "note | spotter | owner | new playbook | triage | null",
      "spotters": [],
      "owner": null,
      "due": null,
      "due_moved": 0,
      "priority": null,
      "outside_judgment": null,
      "heard": null,
      "felt_gap": null,
      "response_gap": null
    }
  ],
  "counts": {
    "rows": 0,
    "raised_by_nobody": 0,
    "felt_higher": 0,
    "felt_lower": 0,
    "felt_matched": 0,
    "responding_below": 0,
    "responding_above": 0,
    "responding_matched": 0,
    "heard": 0
  },
  "verdict": "not assessed | not triaged | no playbook | unowned | sorted"
}