AI Governance Foundations

1

What AI Is, and What It Is Not

How these systems behave

▼

Artificial intelligence, in the forms this policy governs, produces output by predicting likely content from patterns in the data it was trained on. Unless it is connected to a live source, it is not looking anything up, and it has no way to check what it produces against a source of truth.

Fluency is not accuracy. A wrong answer arrives sounding exactly like a right one, and the system gives no signal that anything is off. This is how these tools behave normally. It is not a fault or a sign that something has gone wrong.

No memory of us

An AI system knows nothing about the Company, our fleet, or our procedures unless someone puts that information in front of it. The Company is building a regulatory assistant in Claude, with our manuals, procedures, and company context loaded in, which narrows this gap for the work it covers. Until then, and outside it, assume the tool knows nothing about us.

Models and effort levels

One tool usually offers several models, and often a choice of how long it reasons before answering. A faster model or a quicker mode is not a more reliable one; it has simply done less work. Match the setting to what is at stake, and remember that the standard for checking the output is the same either way.

It fills gaps

Where information is missing, the system supplies plausible content in its place. Plausible is not the same as correct.

Correction does not stick

Correcting a system improves that conversation and does not carry forward to the next one. Standing preferences set in your own user settings do carry across sessions; a correction made inside a conversation does not. If you find yourself correcting the same thing repeatedly, put it in your settings instead.

Output quality follows input quality. General instructions produce general output. Specific instructions, real context, and a clear statement of what correct looks like produce something usable.

Why this matters: these characteristics are the entire reason the Human Ownership Standard exists. An AI system cannot verify its own output, so a named person does. AI Policy §1.1§10.2

2

Recognizing AI Threats

Deepfakes, phishing, prompt injection, social engineering

▼

Threat 1

Deepfakes and synthetic media

Convincing fake audio and video of a real person can be built from a short public sample. An urgent voice or video request to move money, release information, or skip a procedure is treated as unverified until you confirm it through a channel you start, using contact details you already hold.

Threat 2

AI-generated phishing

AI has removed the spelling and phrasing errors that used to make phishing obvious. Well-written no longer means legitimate. Verify the sender, not the writing.

Threat 3

Prompt injection

Instructions can be hidden inside content an AI reads: a web page, a PDF, an email, a document someone sends you. The tool may act on them as though you had given them. The exposure is greatest for tools that browse, read attachments, or act on your behalf.

Threat 4

Social engineering of the AI

A request framed as a test, an emergency, or an authorized exception can produce output the requester was never entitled to. Authorization comes from this policy and the Approved Tool Registry, never from what a prompt claims.

Never call back on a number supplied in the request. Use contact details you already have. A convincing voice is not verification.

Submit a safety report for any of the four. A deepfake or AI-enabled fraud attempt is reportable whether or not it succeeded. AI Policy §10.2§12.4

3

Prompt Hygiene

What never goes into a prompt, in any tool

▼

Everything you type into an AI tool leaves your control the moment you send it. It travels to a vendor, is processed on their systems, and may be retained. Approval does not undo that. An approved tool is one the Company has assessed and accepted for specific data and specific uses; it is not a tool you can put anything into.

So there are two separate questions, and both have to be answered before you paste anything. Is this tool approved for this use? And is this data allowed in a prompt at all?

Never put these into any AI prompt, approved or not. Sensitive Security Information. Passenger or crew personal data. Protected health information. Payment card or bank details. Employee personnel records, disciplinary matters, or compensation. Customer contract terms. Credentials, keys, or tokens of any kind.

The registry entry for each approved tool states what it may be used for and what is prohibited in it. Those prohibitions are the operative limit. Where an entry is silent, the list above still applies.

Describe instead of pasting

Most tasks do not need the sensitive part. Ask for the structure of a letter rather than pasting the letter. Ask how to phrase a finding rather than pasting the finding. You usually need the shape of an answer, not the record itself.

Strip before you send

Remove names, tail numbers, dates, and account numbers before pasting. Replace them with placeholders. Put the real values back yourself afterward, in the document rather than in the tool.

Attachments count

Uploading a file is the same act as pasting its contents. A spreadsheet, a PDF, a screenshot: whatever is in it has been sent. Check what is actually in a document before you attach it, including tabs and columns you were not thinking about.

You cannot unsend it

Deleting a conversation does not retrieve what was already transmitted. If you send something you should not have, that is a reportable data exposure, and reporting it promptly is the whole point of the reporting system.

Two habits make the rest easy. Before you paste, read what you are about to paste, not what you think it is. And if you are weighing whether something is sensitive enough to matter, that hesitation is the answer: leave it out.

Ask Gold Aviation is not an exception. A Company assistant built on an approved tool is still an AI tool, and the same limits apply to what you put into it. AI Policy §10.2§6.6§9.1

4

The Human Ownership Standard

Also known as No AI Slop

▼

The Human Ownership Standard, also referred to as No AI Slop, is the standard this policy applies to every AI-assisted output. Artificial intelligence is a tool, and the individual who uses it is accountable for the result.

Any output produced with AI assistance that carries your name, your team's name, or the Company's name belongs to you, not to the AI system. Output that has not been read, understood, and confirmed accurate is not ready for use.

Sending, publishing, filing, or acting on unreviewed AI output is a violation of this policy. This applies regardless of how the output was generated, which tool produced it, or the time pressure involved.

Ownership of AI-assisted work requires all five of the following:

Read the output in full, not a summary or excerpt.
Independently verify material factual claims.
Edit the output to reflect the Company's voice and the facts as known.
Remove anything that cannot be verified or is not yours to stand behind.
Be able to defend the output if questioned: understanding what it does and why, verifying it behaves as intended, and accepting accountability for the outcome.

Full professional accountability applies to AI-assisted work exactly as it applies to work produced without AI. AI Policy §1.1

5

What the Policy Covers

Scope, exclusions, and the six guiding principles

▼

Covered people. All employees of the Company in every department, including leadership. All contractors, consultants, temporary workers, and agents performing work on behalf of the Company. All third parties who access organizational systems or data while providing services.

Covered systems. Generative AI producing text, images, code, audio, or video. Agentic AI taking autonomous actions. Analytical and decision-support AI embedded in software platforms. AI features within approved applications. Custom or internally developed models.

Not covered: standard spell-check and grammar correction, rule-based automation with no machine learning, and vendor-operated AI where no organizational data is processed and no organizational output is involved. The Human Ownership Standard still applies to everything.

Six principles govern any situation the policy does not explicitly address:

Principle	What it requires
Human ownership	Every AI output that represents or affects the organization has a named human owner before use.
Least privilege	AI gets the minimum access needed for the specific approved task.
Transparency	Disclose AI use where the policy, a contract, a regulation, or professional expectation requires it.
Proportionate risk	Controls match risk. Low-risk use gets light oversight; high-risk use gets rigorous review.
Continuous oversight	Approval today is not approval forever. Tools and access are reviewed on a set cadence.
No silent agents	Autonomous action requires documented authorization, defined scope, logging, and human interrupt.

AI Policy §1.3§1.4§2

6

Review Standards and Disclosure

Minimum review by output type, and when to disclose

▼

Different output types carry different minimum review standards. These supplement, and do not replace, the Human Ownership Standard.

Output type	Minimum review
External communications	Full read and edit by the named sender, who executes the send personally.
Legal or contractual documents	Full review plus legal sign-off. Attorney review for anything legally binding.
Financial analyses and reports	Full review plus numerical verification, with written attestation on file.
Code and technical artifacts	Functional testing by a competent reviewer before use. No autonomous merge or deploy. Deployment requires explicit human action.
Regulatory and compliance filings	Full review plus sign-off by the accountable manager for that filing.
Presentations and board materials	Full review by the presenter, who is accountable for every claim.

No AI-only submission to the FAA, TSA, or any regulator under any circumstances.

Disclosure is required for external documents submitted to regulators, courts, or government agencies including the FAA and TSA; client-facing deliverables where the engagement agreement specifies it; published content where platform or industry rules require it; and internal communications where a material business decision rests on AI-generated analysis.

Standard disclosure language: "This was prepared with AI assistance and reviewed and approved by [NAME], [TITLE]."

7

Consequences and Exceptions

What happens when the standard is not met

▼

Violations are handled consistently and progressively. Deliberate violations, and violations causing material harm, may bypass steps up to and including termination, at the discretion of the President and the AI Steering Committee. This framework supplements existing disciplinary policy rather than replacing it.

Behavior	First	Second	Third
Using an unapproved AI tool	Documented coaching. Access suspended pending review. Intake form required.	Written warning. Access tier downgrade.	Escalation to the President; potential termination.
Distributing unreviewed AI output	Documented coaching. Remedial training within 10 days.	Formal written warning. Manager notified.	Escalation to the President.
Sharing confidential or regulated data in an AI prompt	Incident report filed. Mandatory security training.	Written warning. Access suspension pending review.	Escalation to the President and External Legal.
Non-completion of required training	Access suspended until completion.	Written warning.	Escalation to the President.
Deliberate policy misrepresentation	Immediate escalation to the President and External Legal.

Exceptions. A written request to the CISO stating the provision at issue, the business justification, the duration requested, and the compensating controls. Exceptions cap at 90 days and renew only by re-approval. Every exception is recorded in the Exception Register.

No exception is ever granted to the Human Ownership Standard or the Non-Supersession Rule. Those two provisions are absolute, and a request for an exception to either will be refused. AI Policy §11.4§12.3

Knowledge Check

Acknowledgment