Gains:
- Ability to transform the ethical principles of artificial intelligence use in special education (beneficence, non-maleficence, justice, avoidance of bias) into concrete rules
- Ability to manage the privacy of student and family data, data storage and sharing limits within the framework of KVKK
- Ability to tie learnings throughout the module into an end-to-end secure workflow and define where artificial intelligence ends and expert responsibility begins
Throughout this module, we learned to use AI as an assistant: drafted an IEP goal, adapted materials, established an AAC and behavior plan, monitored progress, and strengthened communication with the family. This final unit looks at the framework that holds it all together: ethics, privacy and expert oversight. Because the "data" you work with in special education is not a table, but a child; The "output" you produce is not a document, but a decision affecting a child's right to education. Artificial intelligence provides great convenience in this field; But when used incorrectly, the people who can be harmed the most are also the most vulnerable. Let's repeat for the last time the principle we established from the beginning: Artificial intelligence is an assistant; Diagnosis, placement, purpose and decisions affecting the child's future belong to the competent specialist, the IEP team and the family.
Four ethical principles and their concrete counterparts
We can attribute the use of artificial intelligence in special education to four basic ethical principles:
- Beneficence: AI should be used to truly contribute to the child's education; It should help the teacher free up his time and pay more attention to the child. Use is not an end in itself.
- Nonmaleficence: Unverified, child-inappropriate, or stigmatizing output can cause harm. Every output passes through the filter of "will this harm the child?"
- Justice: Every child should have equal access to resources and quality education. AI should not privilege some children and neglect others; Everyone should be given individual attention.
- Avoiding bias: AI can reflect stereotypes in the data it is trained on; The specific type of disability may produce discriminatory output about gender, culture, or socioeconomic status. The specialist must recognize and correct these outcomes and evaluate each child individually.
Bias is especially important because it is insidious. The AI can fluently produce a sentence such as “such children usually cannot do X”; However, this is a stereotype that is both wrong and hinders the child. The task of the specialist is to see these patterns and evaluate the child with his potential.
Caution: No AI-generated generalizations ("children with X disabilities can/do Y") can be used to limit a child's purpose. Low expectations are one of the most common and most invisible forms of discrimination.
Privacy: data storage and sharing limits
Student data (identity, diagnosis, RAM report, behavior record, family information) is special personal data within the scope of KVKK. Let's clarify the rules in this unit:
- Anonymize: Name, ID, protocol, school, address and family details are removed before entering any AI tool.
- Minimal data: Share only the information needed for the task; Don't give the entire file "just in case".
- Secure tool: If possible, corporate tools that have a data processing contract and do not use your data in model training are preferred; Sensitive data is not uploaded from personal accounts.
- Storage and deletion: If drafts produced with AI contain student data, they are stored according to institution rules and deleted when necessary; It is not left in random places.
- Sharing: Information is shared only with authorized persons (IEP team, family, relevant specialist) and only as much as necessary.
The table below summarizes the limits learned throughout the module in one place:
area
AI can
AI can't / belongs to humans
Evaluation
Summarizes anonymous data
Diagnoses/changes
BEP
Writes a draft of objectives
Confirms/confirms intent
material
Adapts, differentiates
Makes the curriculum/copyright decision
AAC
Card/story outline
Chooses the method and adapts it to the child
behavior
ABC summarizes, gives advice
Decides the function, approves the plan
progress
Summarizes and interprets data
Final decision to proceed
family
text drafts
Validates text, enforces relationship
The right column in this table is the limit of expert responsibility and cannot be delegated to the AI under any circumstances.
End-to-end secure workflow
Let's combine the module into a single flow. A secure AI-supported tutoring workflow is as follows: (1) anonymize data; (2) write a prompt with clear roles and boundaries; (3) get draft from AI; (4) verify the output in three steps — link to source, compare with child's actual profile, expert-filter; (5) conduct an ethics and bias audit; (6) adapt and validate with team/family; (7) apply and monitor with data; (8) keep/delete drafts according to institution rules. This cycle repeats with each task, keeping clear where AI ends and expert responsibility begins.
three mini cases
Case 1 — Catching bias. A teacher asked the AI for a goal outline; There was a sentence in the output like "this type of students cannot think abstractly, keep the target low." The teacher recognized this as a stereotype; His own student was able to establish abstract relationships. He extracted the sentence and wrote the goal based on the child's true potential. Ethical supervision prevented a low expectation that would hinder the child.
Case 2 — Privacy discipline. One institution made it a rule for all teachers to do an “anonymization check” before using AI. A teacher removed name, protocol, and family information before having a RAM report summarized. Months later, a data audit showed that the organization had not provided any sensitive data to an outside tool. A simple habit protected the organization from serious risk.
Case 3 — Error in delegating responsibility. A teacher, under time pressure, implemented a behavior plan produced by an AI without reading it. The plan did not fit the child's functioning and the behavior escalated. The problem was not the AI's "mistake" but the delegation of responsibility. Unverified output is like an unsigned expert judgment; The responsibility always lies with the expert. Mistake: substituting convenience for responsibility.
Weak prompt / Strong prompt
Weak prompt:
Analyze everything about this student and tell me what to do.
This request both pushes us to share more data (perhaps containing identity) than is necessary by saying "everything" and transfers the decision to the AI. It also violates two basic principles.
Powerful prompt:
Your role: assistant to special education specialist.Use only the following anonymous, task-specific information: [minimum required].Task: Produce a DRAFT for [single, clear task].Rules: Diagnosis/decision making. Do not generalize/stereotype. State that you are not sure. Assume this will be verified by the professional and confirmed with the team/family.
This claim includes minimal data, clear task, bias, and decision limits together; It collects all the principles of the module in a single template.
Common mistakes
- Delegating responsibility. Leaving the decision to the AI; executing without reading the output.
- Not seeing prejudice. Not noticing stereotypical, expectation-lowering generalizations.
- Sharing excessive data. Giving unnecessary, sensitive data to the tool by saying "analyze everything".
- Bypassing anonymization. Working without clearing identity, diagnosis and family information.
- Storing drafts insecurely. Leaving printouts containing student data in random places.
- Seeing ethical auditing as formal. Skipping questions like “will it cause harm, is it fair, is it biased?”
In summary
In special education, AI is a powerful assistant, but because you are working with the most vulnerable individuals, ethics and privacy are non-negotiable. Four principles guide us: beneficence, nonmaleficence, justice, avoidance of bias. Student data is anonymized, shared with minimal and secure means, and stored securely. End-to-end secure workflow; The cycle is anonymize, get draft, verify in three steps, ethical review, approve with team/family, track with data. Wherever AI ends in this cycle, expert responsibility begins. An unverified output is like an unsigned expert judgment; The signature is always yours.
Application task
Choose one of the outlines you produced throughout the module (IEP goal, material, behavior plan, or family text) and take it through all of the "end-to-end secure workflow" steps in this unit. In particular, check for ethics/bias: is there a lowering generalization, a stigmatizing statement, or a privacy leak in the output? Fix any issues you find and note which step of the flow caught that issue.
checklist
- [ ] I anonymized the data; I only shared the necessary minimum.
- [ ] I verified the output with three steps (source/profile/expert).
- [ ] I cleared the prejudices and expectations-lowering generalizations.
- [ ] I weeded out stigmatizing language; I observed four ethical principles.
- [ ] I kept the drafts safe/deleted them when necessary.
- [ ] I took the final decision and responsibility as the expert/team/family.
Module Exam
1. A special education teacher asks the artificial intelligence 'what is the diagnosis of this child' by simply giving the student's name and age and writes the answer into the IEP file. What is the fundamental mistake in this approach?
- A) Ignoring that the authority to make a diagnosis belongs to RAM and authorized experts; also move an unverified output to an official record ✔
- B) It is unnecessary to give age information to artificial intelligence
- C) It is forbidden to keep the BEP file in digital environment
- D) Artificial intelligence always makes the diagnosis seem heavier than it is
Explanation: Artificial intelligence cannot and should not diagnose; Diagnosis and evaluation are the job of RAM and authorized experts. In addition, without giving data to the model, the desired output becomes a fabrication (hallucination) and sharing the student identity is a violation of privacy.
2. Which of the following is the main feature that a short-term (measurable) BEP goal should have?
- A) Aiming for the student to generally like the subject
- B) It is based only on the teacher's observation and does not contain numerical criteria
- C) Gather as long and as many outcomes as possible in one sentence.
- D) Contains observable behavior, condition and a measurable success criterion ✔
Explanation: The short-term goal includes an observable behavior, condition, and criterion (how many trials, what percentage, what level of independence). Immeasurable expressions such as 'understands', 'learns', 'becomes aware of' cause errors in writing the purpose.
3. A teacher uploads the RAM report, which includes the student's full name, TR ID number and diagnosis, to a publicly available artificial intelligence tool and says 'summarize'. What would be the right approach?
- A) Upload the report without any modifications, because the summary requires all the information
- B) Just delete the TR ID number and leave the name and diagnosis
- C) Removing identity information, anonymizing data and choosing a secure/institutional tool ✔
- D) Sharing the report via e-mail from your personal account and summarizing it accordingly
Explanation: Student data is special personal data within the scope of KVKK. Identity information should be removed and the data anonymized ('10-year-old male student'), and if possible, a secure tool that is institutional and does not use the data in education should be preferred.
4. What is the best attitude when creating a social story with artificial intelligence for alternative and augmentative communication (AAC)?
- A) Adapt the draft specifically for the child and have it approved by an expert ✔
- B) Using the produced story exactly for all students without changing it at all
- C) Writing the story with as long and abstract words as possible
- D) Giving only long text without adding visual support
Description: Artificial intelligence generates drafts; However, social history and communication support should be adapted to the child's individual characteristics and approved by the speech-language pathologist/specialist. A general outline cannot be directly applied.
5. Which of the following is the basic principle when getting help from artificial intelligence in a positive behavior support plan?
- A) Asking him/her to recommend the harshest punishment to quickly suppress the behavior.
- B) Understand the function of behavior and produce an alternative positive behavior and preventive strategy draft ✔
- C) Finalizing the plan without knowing the student, just looking at the name of the behavior
- D) Implementing the plan without involving the family and specialist in the process
Explanation: Positive behavior support is not punishment-oriented; It aims to understand the function of behavior (functional evaluation), preventive regulations and provide a teachable alternative behavior instead of undesirable behavior. Expert supervision is essential.
6. What does 'data-driven decision' mean in tracking progress?
- A) Making a decision based only on a single observation at the end of the year
- B) Writing a progress report based on the teacher's general impression without looking at the data
- C) Leaving the decision entirely to the interpretation of artificial intelligence
- D) Making a decision to maintain, adapt or change the goal by looking at regular measurement data ✔
Explanation: Data-based decision making is deciding to maintain, adapt or change the goal by looking at regular achievement-based measurements (percentage, number of attempts, level of independence). Artificial intelligence summarizes data and suggests interpretation; The expert makes the decision.
7. According to the universal design (UDL) principle, what is the best way to make a material accessible?
- A) Presenting the material in only one format, in small print and dense text
- B) Presenting content in multiple ways, providing plain language and alternative text ✔
- C) Remove all adaptations and give everyone the same standard material
- D) Assuming accessibility is achieved only by changing color
Description: UDL envisions presenting content in multiple ways (text, visual, audio), plain language, alternative text and flexible formats. The aim is to design to suit different needs, not just one 'average' student. Final inspection depends on expert and real user testing.
8. A teacher sends the family information text produced by artificial intelligence to the family without checking it, and the text contains an incorrect appointment date and an exaggerated 'recovery' promise. What is the fundamental mistake here?
- A) Creating false information and unrealistic expectations by sending the text without verifying it ✔
- B) Sending a written text to the family is itself wrong.
- C) The text is too short
- D) Artificial intelligence is strictly prohibited from producing family text
Description: Every text shared with the family must be produced with accuracy, confidentiality and expert approval. Artificial intelligence is fluent but can produce false information (hallucinations); Text sent without verification damages trust and process.
9. Which of the following is the correct approach when making differentiation in an inclusion/integration class?
- A) Separating the student completely from the classroom and having him/her do each activity alone
- B) Supporting the same achievement with shared responsibility, without adapting and labeling it according to the level ✔
- C) Giving the same material to the entire class without exception and not making any adaptations
- D) Leaving the adaptation only to the special education teacher and removing the classroom teacher from the process
Explanation: Differentiation means adapting the same learning outcome to different levels and supporting the student without labeling them. Adaptation is the joint responsibility of the classroom teacher and the special education teacher; Solutions that separate and stigmatize the student from their peers are avoided.
10. Why is it particularly risky for AI language models to produce 'hallucinations' in the context of special education?
- A) Because the model always runs slowly
- B) Because the model does not know Turkish
- C) Because hallucination occurs only in visual production
- D) Because made-up information can affect the child's educational decision and can be assumed to be true in fluent language ✔
Explanation: A hallucination is when the model fluently fabricates non-existent information as if it were true. A fabricated piece of legislation, an incorrect development norm or an imaginary method in special education may directly affect the child's right to education and development; so every output must be validated.
11. What does 'avoiding bias' mean in the ethical principles of using artificial intelligence in special education?
- A) Recognizing and correcting stereotypical or discriminatory outputs produced by the model and evaluating each child individually ✔
- B) Not writing long prompts just to get quick output.
- C) Using artificial intelligence on only one type of obstacle
- D) Applying the same biased template to all students
Explanation: Artificial intelligence may reflect biases in the data on which it is trained; may produce stereotypical output about the particular type of disability, gender, socioeconomic status, or culture. The specialist should monitor the outcome in this respect and evaluate each child individually.
12. Why is it important to include a 'strengths' section when summarizing a RAM report with artificial intelligence and creating a student profile?
- A) To make the report appear longer
- B) Strengths are used only for form, although they are not legally required.
- C) To base planning on strengths and not define the student only by his/her shortcomings ✔
- D) To completely ignore the needs and write only positively
Explanation: Planning in special education is based on strengths; Only a deficiency/need-oriented profile defines the student with his/her shortcomings. Strengths are the basis for goals and motivation and provide a positive, growth-oriented outlook.
13. What is the correct attitude in terms of 'copyright and originality' in the production of adapted material?
- A) Reproducing the output with copyrighted characters and content without checking
- B) Checking and originalizing the material in terms of compliance with the curriculum, accuracy and copyright ✔
- C) Distribute every image commercially, regardless of the source
- D) Assuming that the copyright issue does not apply to digital material at all
Explanation: Artificial intelligence output may have been fed from other sources; may contain copyrighted content, brand characters or unverified information. The expert should check the material for compliance with the curriculum, accuracy and copyright, and cite the source and originalize it when necessary.
14. How does the key principle emphasized throughout the module describe the role of artificial intelligence in special education?
- A) Artificial intelligence replaces the expert and makes decisions automatically
- B) Artificial intelligence can only be used for note-taking and nothing else
- C) When artificial intelligence is used, expert supervision is no longer necessary
- D) Artificial intelligence is an assistant and draft generator; The final decision and responsibility lies with the specialist, the team and the family ✔
Description: Artificial intelligence is an assistant and sketch generator; Diagnosis, placement, goal setting and decisions affecting the child's future belong to the competent specialist, the IEP team and the family. Unverified output is like an unsigned expert verdict.